RU2810934C2 - Method for training machine learning algorithm for predicting intent parameter for object on ground for use in unmanned vehicles - Google Patents

Abstract

FIELD: machine learning.

SUBSTANCE: group of inventions is related to a method for training a machine learning algorithm to predict the intent parameter for an object on the ground for use in unmanned vehicles and a method for operating an unmanned vehicle. An unmanned vehicle moves through the terrain near the object. The method is comprised of allowing the server to access operational data associated with the training vehicle; generation by the server of a training set for training the machine learning algorithm. The operating data includes data indicating the training location, the state of the training vehicle, and the state of the training object at the training location. The training set has input information and an accessory-free label. The input information includes data indicating the training location and target states of the training vehicle and training object. The accessory-free label is based on data indicating the future states of the training vehicle and training object at the training location at the target time.

EFFECT: improved performance of unmanned or semi-autonomous vehicles is achieved.

20 cl, 8 dwg

Description

TECHNICAL FIELD

[0001] The present technology relates to computer-implemented methods for training machine learning algorithms, and more particularly, to methods for generating and training a machine learning algorithm to predict an intent parameter for an object in the field for use in self-driving vehicles.

BACKGROUND OF THE ART

[0002] Self-driving vehicles are vehicles that are capable of autonomously navigating private and/or public spaces. Using a system of sensors that detect the location and/or environment of the vehicle, logic within or associated with the vehicle controls the speed, traction, braking, and steering of the vehicle based on the location and environment of the autonomous vehicle detected by the sensors.

[0003] A self-driving vehicle may utilize a variety of sensor systems, such as, but not limited to, camera systems, radar systems, and light detection and ranging (LIDAR) systems. Different sensor systems can be used to capture different information, and/or in a different format, about the location and environment of the autonomous vehicle. For example, camera systems can be used to capture image data of a driverless vehicle's surroundings. In another example, LIDAR systems can be used to capture point cloud data to construct 3D map representations of the environment and other potential objects located in the vicinity of the autonomous vehicle.

[0004] One challenge associated with properly operating an unmanned vehicle is correctly determining the likely intent of other moving objects, such as other vehicles, in the vicinity of the unmanned vehicle. This determination can be influenced by a large number of factors, including detected information about the configuration and movement of surrounding objects, the movement of the autonomous vehicle, and the terrain over which the autonomous vehicle and other objects are moving. If this determination is not correct, it may result in sudden maneuvers or even a collision.

[0005] While various predictive analytics approaches can be used to make this determination, many of them, such as machine learning algorithms, must be trained using sets of pre-labeled training data. Training a machine learning algorithm to accurately predict the “intent” of a moving object from sensor data available to an autonomous vehicle may require large sets of pre-labeled training data, which can be time-consuming and expensive to generate.

[0006] US Patent 10,739,775 discloses a system and method for simulating the trajectory of an autonomous vehicle in the real world. Receive training data from a real world data acquisition system, obtain ground truth data, and train multiple trajectory prediction models. The simulation phase is performed to generate the closest scenario for each vehicle being simulated. Nearby scenarios correspond to different locations, traffic patterns, or environmental conditions. Vehicle intent data corresponding to representing various types of simulated driver or vehicle intent data may be provided.

SUMMARY OF THE INVENTION

[0007] As discussed above, training a machine learning algorithm to reliably predict the intent of a moving object may require the use of a large training data set. A “naive” approach to creating a training dataset is to collect data about the state of the autonomous vehicle, the state of the object and the terrain, and manually label each element of such data using an intent parameter. Manually generating large datasets to train machine learning algorithms is extremely expensive and time-consuming.

[0008] Accordingly, the present technology provides an alternative to using manually labeled data sets to train machine learning algorithms to determine the intent parameter on the ground. Specifically, the present technology uses an automated approach to generate large training data sets to train machine learning algorithms for use with autonomous vehicles to determine the intent parameter. The technology then uses machine learning algorithms that were trained using these data sets to operate the self-driving vehicle. Accordingly, the present technology effectively improves the performance of unmanned or semi-autonomous vehicles.

[0009] In accordance with one aspect of the present disclosure, the technology is implemented in a method of training a machine learning algorithm to predict an intent parameter for an object in a landscape in which an unmanned vehicle is moving through that terrain in close proximity to the object. The method is executable by the server and includes the server accessing operational data associated with the training vehicle, the operational data having previously been stored in storage. The operating data includes data indicating: the training location, the state of the training vehicle at the training location, and the state of the training object at the training location.

[00010] For a target time point in the operation data, the method includes retrieving from the operation data by a server data indicating a training location at the target time point, a target state of the training vehicle corresponding to the target time point, and a future state of the training vehicle after the target time point time, the target state of the learning object corresponding to the target time, and the future state of the learning object after the target time. The method also includes the server generating a training set for training the machine learning algorithm, wherein the training set has input information and an assessor-less label. The input information includes data indicating the training location and target states of the training vehicle and training object. An accessory-free label is based on data indicating the future states of the learning vehicle and learning object. An accessory-free label indicates the intent of the training object at the training location at the target time.

[00011] The method further includes training, by the server, a machine learning algorithm based on the training set to predict, during the use phase, an intent parameter indicating the intent of the terrain object at the target time of use, based on data indicating (i) the terrain at the target use time point, (ii) the use state of the unmanned vehicle at the target use time point, and (iii) the use state of the said object.

[00012] In some implementations, retrieval includes the server retrieving data indicating (i) a set of previous states of the training vehicle prior to a target time point and (ii) a set of previous states of the training vehicle prior to the target time point. The input information further includes data indicating sets of previous states of the training vehicle and the training object.

[00013] In some implementations, the set of previous states of the training vehicle is a portion of the training vehicle's trajectory to a target time, wherein the portion of the trajectory corresponds to a time interval of a predetermined duration to the target time. In some implementations, the set of previous states of the training object is part of the trajectory of the training object to the target time, and part of the trajectory corresponds to a time interval of a predetermined duration to the target time.

[00014] In some implementations, the retrieval includes the server retrieving data indicating (i) a set of future states of the training vehicle after the target time, including said future state of the training vehicle, and (ii) a set of future states of the training object after the target point in time, including said future state of the learning object. The accessory-free label in these implementations can be based on sets of future states of the learning vehicle and the learning object. In some of these implementations, a set of future states of the training vehicle is part of the training vehicle's trajectory after the target time, and a set of future states of the training object is part of another trajectory of the training object after the target time.

[00015] In some implementations, the method further includes the server selecting a target time point in the operational data. The selection includes the server identifying a certain state of the training vehicle and a certain state of the training object in the operating data that are associated with the same position on the training terrain, and selecting by the server a target time point as a previous point in time relative to a certain point in time associated with a certain state of the training vehicle.

[00016] In some implementations, the method includes the server generating an accessoryless label based on data indicating future states of the training vehicle and the training object. Generating an accessoryless label includes determining by the server that the future states of the training vehicle and the training object indicate at least one of: the training vehicle and the training object, after the target time, occupy the same position on the training terrain, and that the training vehicle occupies that one position before the training object; the training vehicle and the training object occupy one position on the training terrain after the target time and that the training object occupies this one position before the training vehicle; the training vehicle and the training object do not occupy the same position on the training area after the target time; and that the training vehicle occupies a certain position on the training terrain at a future time point, and that the training vehicle occupies another specific position on the training terrain at a future time point, and the distance between said specific position and the other specific position is below a threshold safe distance.

[00017] In some implementations, the method further includes the server transmitting information representing the machine learning algorithm to an electronic device associated with the autonomous vehicle. In these implementations, at a target point in time of use, during operation of the autonomous vehicle, the method includes executing a machine learning algorithm by the electronic device to generate an intent parameter, and causing the electronic device to operate the autonomous vehicle at least in part based on the intent parameter.

[00018] In some implementations, the machine learning algorithm is configured to classify the intent of an object, wherein the intent parameter indicates the likelihood that the object's intent is one of a plurality of intent classes. In some implementations, the training vehicle is an unmanned vehicle. In some implementations, the object is a moving object. In some implementations, the object is at least one of a vehicle and a pedestrian.

[00019] In some implementations, the training terrain includes a portion of the training road. The training location data may be a bird's eye view image of a section of the training road at a target time.

[00020] In accordance with other aspects of the present disclosure, a method of operating an unmanned vehicle is provided. The method includes receiving, at an electronic device associated with the unmanned vehicle, data from a sensor associated with the unmanned vehicle, the data indicating the state of the unmanned vehicle, the state of an object in the vicinity of the unmanned vehicle, and the terrain over which the unmanned vehicle is moving. The method further includes providing said data to a machine learning algorithm on the electronic device, wherein the machine learning algorithm determines an intent parameter indicating the intent of an object on the ground at the time of the target time, and causing the electronic device to operate the unmanned vehicle at least in part based on the parameter intentions. The machine learning algorithm was trained using a training data set generated by, at least in part: a server accessing operational data associated with a training vehicle, the operational data having previously been stored in storage and including data indicating: the training location , the state of the training vehicle at the training site and the state of the training object at the training site. Generating the training set further included, for the target training time in the operating data, retrieving from the operating data via a server data indicating: the training location at the target training time, the target state of the training vehicle corresponding to the target training time, and the future state the training vehicle after the target training time; and a target state of the learning object corresponding to the target learning time, and a future state of the learning object after the target learning time. Generating the training set further included the server generating a training set of data for training the machine learning algorithm, wherein the training data set has input information and an accessory-free label, wherein the input information includes data indicating the training terrain and target states of the training vehicle and the training object, and the accessory-free label is based on data indicating the future states of the learning vehicle and the learning object. An accessory-free label indicates the intent of the training object at the training location at the target training time.

[00021] In some implementations of this aspect, the machine learning algorithm is configured to classify the intent of an object, wherein the intent parameter indicates the likelihood that the object's intent is one of a plurality of intent classes. In some implementations of this aspect, the method further includes receiving, on the electronic device, information representing the machine learning algorithm from the server.

[00022] In accordance with a further aspect of the present disclosure, this technology is implemented in a self-driving vehicle. The unmanned vehicle includes at least one sensor configured to collect sensor data indicating the state of the unmanned vehicle, the state of an object in the vicinity of the unmanned vehicle, and the terrain over which the unmanned vehicle is moving. The unmanned vehicle further includes an electronic device including a processor and memory. The memory stores programmed instructions that, when executed by the processor, cause the electronic device to: provide sensor data to a machine learning algorithm running on the processor, wherein the machine learning algorithm determines an intent parameter indicating the intent of an object on the ground at the time of the target time; and initiate operation of the autonomous vehicle based at least in part on this intent parameter. The machine learning algorithm includes values stored in memory that were determined by training the machine learning algorithm using a training data set generated by, at least in part: a server accessing operational data associated with the training vehicle, wherein the operational data were previously stored in storage and include data indicating: the training location, the state of the training vehicle at the training location, and the state of the training object at the training location. Generating the training set further included, for the target training time in the operating data, retrieving from the operating data via a server data indicating: the training location at the target training time, the target state of the training vehicle corresponding to the target training time, and the future state the training vehicle after the target training time; and a target state of the learning object corresponding to the target learning time, and a future state of the learning object after the target learning time. Generating the training set further included the server generating a training set of data for training the machine learning algorithm, wherein the training data set has input information and an accessory-free label, wherein the input information includes data indicating the training terrain and target states of the training vehicle and the training object, and the accessory-free label is based on data indicating the future states of the learning vehicle and the learning object. An accessory-free label indicates the intent of the training object at the training location at the target training time.

[00023] In some implementations according to this aspect of this disclosure, the machine learning algorithm is configured to classify the intent of an object, wherein the intent parameter indicates the likelihood that the object's intent is one of a plurality of intent classes. In some implementations of this aspect, the memory further stores programmed instructions that, when executed by the processor, cause the electronic device to receive information representing the machine learning algorithm from the server.

[00024] Additional and/or alternative features, aspects and advantages of implementations of the present technology will become apparent from the following description, the accompanying drawings and the appended claims.

BRIEF DESCRIPTION OF THE DRAWINGS

[00025] These and other features, aspects and advantages of the present technology will become more apparent from the following description, the appended claims and the accompanying drawings, in which:

[00026] FIG. 1 illustrates a schematic representation of an exemplary computer system for use in certain implementations of systems and/or methods of the present technology.

[00027] FIG. 2 illustrates a networked computing environment including an autonomous vehicle suitable for use with certain implementations of the present technology.

[00028] FIG. 3 shows a timeline of the data collection process through the learning vehicle.

[00029] FIG. 4 illustrates an example scenario of data collection by a training vehicle at a training location.

[00030] FIG. 5 illustrates the process of training a machine learning algorithm using operational data collected by the training vehicle.

[00031] FIG. 6 shows an example scenario for a machine learning algorithm used in a self-driving vehicle.

[00032] FIG. 7 depicts the generation of a predictive intent parameter for an object by a trained machine learning algorithm using in-use data.

[00033] FIG. 8 illustrates a flowchart of a method for training a machine learning algorithm to predict a terrain object's intent parameter and using the trained machine learning algorithm.

DETAILED DESCRIPTION

[00034] Various representative implementations of the disclosed technology will be described in more detail below with reference to the accompanying drawings. However, the present technology can be implemented in many different forms and should not be construed as limited to the exemplary implementations disclosed herein. In the drawings, the dimensions and relative sizes of layers and regions may be enlarged for clarity. Everywhere the same numbers refer to the same elements.

[00035] The examples and conditional statements provided herein are intended primarily to assist the reader in understanding the principles of the present technology and not to limit its scope to such specific examples and conditions. It should be understood that those skilled in the art will be able to develop various mechanisms that, although not explicitly described herein, nevertheless embody the principles of the present technology and are included in the spirit and scope thereof.

[00036] In addition, the following description may describe implementations of the present technology in a relatively simplified form for purposes of ease of understanding. Those skilled in the art will appreciate that various implementations of the present technology can be highly complex.

[00037] In some cases, examples of modifications to the present technology that are considered useful may also be set forth. This is intended only to facilitate understanding and, again, not to strictly define the scope or delineate the boundaries of the present technology. These modifications are not an exhaustive list, and one skilled in the art will be able to make other modifications while still remaining within the scope of the present technology. In addition, where examples of modifications are not provided, it should not be construed to mean that no modifications can be made and/or that what is described is the only way to implement such element of the present technology.

[00038] It should be understood that although the terms first, second, third, etc. may be used in this document to describe various elements, these elements should not be limited to these terms. These terms are used to distinguish one element from another. Thus, the first element discussed below could be referred to as the second element without departing from the teachings of the present disclosure. As used herein, the term “and/or” includes any and all combinations of one or more related listed elements.

[00039] It should be understood that when an element is referred to as "connected" or "linked" to another element, it may be directly connected or coupled to the other element, or intermediate elements may be present. In contrast, when an element is referred to as "directly connected" or "directly connected" to another element, no intermediate elements are present. Other words used to describe the relationship between elements should be interpreted in a similar way (for example, “between” or “immediately between”, “adjacent” or “immediately adjacent”, etc.).

[00040] The terminology used herein is intended to describe specific representative implementations only and is not intended to limit the present technology. The singular forms used herein are intended to include the plural forms unless the context clearly indicates otherwise. In addition, it will be understood that the terms “comprises” and/or “comprising,” when used in this specification, define the presence of the claimed functions, entities, steps, operations, elements and/or components, but do not exclude the presence or addition of one or more other functions, wholes, stages, operations, elements, components and/or groups thereof.

[00041] The functions of the various elements shown in the figures, including any functional block labeled "processor", may be provided through the use of dedicated hardware as well as hardware capable of executing software. When provided by a processor, the functions may be provided by one dedicated processor, one shared processor, or a plurality of separate processors, some of which may be shared. In some implementations of the present technology, the processor may be a general purpose processor, such as a central processing unit (CPU), or a purpose-specific processor, such as a digital signal processor (DSP). In addition, explicit use of the term "processor" should not be construed as referring solely to hardware capable of executing software, and may implicitly include, but is not limited to, application-specific integrated circuit (ASIC), field programmable gate array (FPGA), constant memory (ROM) for storing software, random access memory (RAM) and long-term storage. Other hardware, traditional and/or custom, may also be included.

[00042] Software modules, or simply modules or blocks, which are contemplated to be software, may be represented herein as any combination of flowchart elements or other elements indicating the execution of process steps and/or textual description. Such modules may be executed by hardware, which is shown explicitly or implicitly. It is further understood that a module may include, for example, but not limited to, computer program logic, computer program instructions, software, a stack, firmware, hardware circuitry, or a combination thereof that provides the required capabilities.

[00043] As used herein, a “database” is any structured collection of data, regardless of its specific structure, database administration software, or computer hardware on which data is stored, implemented, or otherwise made available for use. . The database may reside on the same hardware as the process that stores or uses the information stored in the database, or it may reside on separate hardware, such as a dedicated server or multiple servers.

[00044] The present technology may be implemented as a system, method, and/or computer program product. The computer program product may include a computer-readable storage medium (or media) storing computer-readable program instructions that, when executed by a processor, cause the processor to implement aspects of the disclosed technology. The computer-readable storage medium may be, for example, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. A non-exhaustive list of more specific examples of computer readable storage media includes: portable computer disk, hard disk, random access memory (RAM), read only memory (ROM), flash memory, optical disk, memory card, floppy disk, mechanically or visually encoded media (for example, a punched card or a barcode) and/or any combination thereof. The computer-readable media used in this document should be considered a non-transitory computer-readable media. It should not be considered a transient signal such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating along a waveguide or other transmission media (such as light pulses traveling along a fiber optic cable), or electrical signals transmitted along a wire.

[00045] It will be appreciated that computer readable program instructions may be downloaded to appropriate computing or processing devices from a computer readable storage medium or to an external computer or external storage device via a network such as the Internet, a local area network, a wide area network, and/or a wireless network. . A network interface on a computing/processing device may receive computer-readable program instructions via a network and forward computer-readable program instructions for storage on a computer-readable storage medium in an associated computing or processing device.

[00046] Computer-readable program instructions for performing the operations of the present disclosure may be assembly instructions, machine instructions, firmware instructions, configuration data for an integrated circuit, or either source code or object code written in any combination of one or more programming languages. Computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network.

[00047] All statements in this document setting forth the principles, aspects and implementations of this technology, as well as specific examples thereof, are intended to cover both their structural and functional equivalents, whether currently known or to be developed in the future. future. Thus, for example, those skilled in the art will recognize that any flowcharts herein represent conceptual views of an illustrative circuit embodying the principles of the present technology. Likewise, it will be understood that any flowcharts, flowcharts, state-change diagrams, pseudo-codes, and the like represent various processes that may inherently be represented in computer-readable program instructions. These computer readable program instructions may be provided to a processor or other programmable processing device to generate a mechanism such that the instructions, which are executed by the computer processor or other programmable processing device, form a means for implementing the functions/actions defined in the sequence block diagram operations and/or flowchart, block by block. These computer readable program instructions may also be stored on a computer readable storage medium that can control a computer, programmable data processing device, and/or other devices to operate in a particular manner, such that the computer readable storage medium having the instructions stored thereon comprises a product of manufacture including instructions that implement aspects of a function/action defined in flowcharts, flowcharts, state-change diagrams, pseudocode, and the like.

[00048] Computer readable program instructions may also be loaded into a computer, other programmable data processing device, or other devices to cause a series of functional steps to be executed on the computer, other programmable device, or other devices to generate a computer-implemented process such that the instructions that are executed on a computer, other programmable device, or other device, implement functions/actions defined in flowcharts, flowcharts, state-change diagrams, pseudocode, and the like.

[00049] In some alternative implementations, functions noted in flowcharts, flowcharts, state diagrams, pseudocode, and the like may not occur in the order shown in the figures. For example, two blocks shown sequentially in a flowchart may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order, depending on the functionality involved. It should also be noted that each of the functions noted in the figures, and combinations of such functions, may be implemented by specialized hardware systems that perform the indicated functions or actions, or by combinations of specialized hardware and computer instructions.

[00050] Having established these basic principles, we now consider some non-limiting examples to illustrate various implementations of aspects of the present disclosure.

Computer system

[00051] FIG. 1 illustrates a computer system 100. The computer system 100 may be a multi-user computer, a single-user computer, a laptop computer, a tablet computer, a smartphone, an embedded control system, or any other computer system now known or later developed. Additionally, it should be understood that some or all of the components of the computer system 100 may be virtualized and/or cloud-based. As shown in FIG. 1, computer system 100 includes one or more processors 102, memory 110, storage interface 120, and network interface 140. These system components are interconnected via a bus 150, which may include one or more internal and/or external buses ( shown) (e.g., PCI bus, Universal Serial Bus, IEEE 1394 Firewire bus, SCSI bus, Serial-ATA bus, etc.) to which various hardware components are electronically coupled.

[00052] Memory 110, which may be random access memory or any other type of memory, may contain data 112, operating system 114, and program 116. Data 112 may be any data that serves as input to or output from any program. in the computer system 100. The operating system 114 is an operating system such as MICROSOFT WINDOWS or LINUX. Program 116 may be any program or set of programs that includes programmed instructions that can be executed by a processor to control actions taken by computer system 100.

[00053] Storage interface 120 is used to connect storage devices, such as storage device 125, to computer system 100. One type of storage device 125 is a solid-state drive, which may use an integrated circuit assembly to permanently store data. Another type of storage device 125 is a hard disk drive, such as an electromechanical device that uses magnetic storage to store and retrieve digital data. Similarly, storage device 125 may be an optical drive, a card reader that accepts a removable memory card, such as an SD card, or a flash memory device that may be connected to the computer system 100, such as via a universal serial bus (USB). .

[00054] In some implementations, the computer system 100 may use well-known virtual memory techniques that allow programs of the computer system 100 to behave as if they had access to a large, contiguous address space instead of accessing multiple, smaller storage spaces such as memory 110 and storage device 125. Therefore, although data 112, operating system 114, and programs 116 are shown to reside in memory 110, those skilled in the art will appreciate that these elements are not necessarily entirely contained in memory 110 at the same time. time.

[00055] Processors 102 may include one or more microprocessors and/or other integrated circuits. Processors 102 execute software instructions stored in memory 110. When computer system 100 starts up, processors 102 may initially execute boot routines and/or software instructions comprising operating system 114.

[00056] Network interface 140 is used to connect computer system 100 to other computer systems or network devices (not shown) via network 160. Network interface 140 may include a combination of hardware and software to enable communication over network 160. In some implementations network interface 140 may be a wireless network interface. Software in network interface 140 may include software that uses one or more network protocols to communicate over network 160. For example, network protocols may include TCP/IP (Transmission Control Protocol/Internet Protocol).

[00057] It should be understood that the computer system 100 is an example only, and that the disclosed technology may be used with computer systems or other computing devices having other configurations.

Network computing environment

[00058] FIG. 2 illustrates a networked computing environment 200 suitable for use with certain non-limiting embodiments of the present technology. Networked computing environment 200 includes an electronic device 210 associated with a vehicle 220 and/or associated with a user (not shown) who is associated with the vehicle 220 (eg, an operator of the vehicle 220). The environment 200 also includes a server 235 communicating with the electronic device 210 via a communications network 240 (eg, the Internet or the like).

[00059] In at least some non-limiting embodiments of the present technology, electronic device 210 is communicatively coupled to control systems of vehicle 220. Electronic device 210 may be located or configured to control various operating systems of vehicle 220, including, but not limited to : ECU (engine control unit), steering systems, brake systems, and alarm and lighting systems (i.e. headlights, brake lights, and/or turn signals). In such an implementation, vehicle 220 may be an autonomous vehicle 220.

[00060] In some non-limiting implementations of the present technology, network computing environment 200 may include a GPS satellite (not shown) transmitting and/or receiving a GPS signal to/from electronic device 210. It should be understood that the present technology is not limited to GPS and may use positioning technology other than GPS. It should be noted that the GPS satellite may not be used at all.

[00061] The vehicle 220 to which the electronic device 210 is coupled may be any recreational or otherwise vehicle, such as a private or commercial automobile, truck, motorcycle, or the like. Although vehicle 220 is illustrated as a ground vehicle, this is not required in every non-limiting implementation of the present technology. For example, in some non-limiting implementations of the present technology, vehicle 220 may be a watercraft, such as a boat, or an aircraft, such as a flying drone.

[00062] Vehicle 220 may be user driven or a driverless vehicle. In some non-limiting implementations of the present technology, it is contemplated that vehicle 220 may be implemented as a self-driving car (SDC). Autonomous vehicle 220 may be a fully autonomous vehicle or a partially autonomous vehicle in which a human driver controls some aspects of the vehicle's operation while other aspects (such as "autopilot" on highways) are automated. It should be noted that the specific parameters of the vehicle 220 are not limiting, these specific parameters include, for example: vehicle manufacturer, vehicle model, vehicle year, vehicle weight, vehicle dimensions, vehicle weight distribution, surface area vehicle, vehicle height, transmission type (such as 2x or 4x), tire type, brake system, fuel system, mileage, vehicle identification number and engine size, or other vehicle characteristics or parameters.

[00063] According to the present technology, the implementation of the electronic device 210 is not particularly limited. For example, electronic device 210 may be implemented as a vehicle engine control unit, a vehicle CPU, a vehicle navigation device (eg, TomTom™, Garmin™), a tablet, a personal computer built into the vehicle 220, and the like. Thus, it should be noted that the electronic device 210 may or may not be permanently coupled to the vehicle 220. Additionally or alternatively, the electronic device 210 may be implemented in a wireless communication device such as a mobile phone (eg, a smartphone or cordless phone). In some implementations, electronic device 210 has a display 270.

[00064] Electronic device 210 may include some or all of the components of the computer system 100 illustrated in Figure 1, depending on the particular implementation. In some implementations, electronic device 210 is an on-board computer device and includes processors 102, storage 125, and memory 110. In other words, electronic device 210 includes hardware and/or software and/or firmware, or both combination for data processing, as will be described in more detail below.

[00065] In some non-limiting implementations of the present technology, communications network 240 is the Internet. In alternative, non-limiting implementations of the present technology, communications network 240 may be implemented as any suitable local area network (LAN), wide area network (WAN), private communications network, or the like. It should be clearly understood that implementations for communication network 240 are provided for illustrative purposes only. A communication link (not specifically numbered) is provided between the electronic device 210 and the communication network 240, the implementation of which will depend, among other things , on how the electronic device 210 is implemented. By way of example, and not limitation, in those non-limiting implementations of the present technology, where the electronic device 210 is implemented as a wireless communication device, such as a smartphone or navigation device, the communication link may be implemented as a wireless communication link. Examples of wireless communication links may include, but are not limited to, a 3G communication network link, a 4G communication network link and the like. Communications network 240 may also use a wireless connection to server 235.

[00066] In some implementations of the present technology, server 235 is implemented as a computer server and may include some or all of the components of the computer system 100 shown in Figure 1. In one non-limiting example, server 235 is implemented as a Dell™ PowerEdge™ server running an operating system Microsoft™ Windows Server™ system, but may also be implemented in any other suitable hardware, software and/or firmware, or combination thereof. In the illustrated non-limiting implementations of the present technology, server 235 is the only server. In alternative, non-limiting implementations of the present technology, the functionality of server 235 may be distributed and implemented by multiple servers (not shown).

[00067] In some non-limiting implementations of the present technology, processors 102 of electronic device 210 may communicate with server 235 to receive one or more updates. Such updates may include, but are not limited to, software updates, map updates, route updates, weather updates and the like. In some non-limiting implementations of the present technology, the electronic device 210 may also be configured to transmit to the server 235 some operational data, such as routes traveled, traffic data, performance data, and the like. Some or all of such data transmitted between vehicle 220 and server 235 may be encrypted and/or anonymized.

[00068] It should be noted that the electronic device 210 may use various sensors and systems to collect information about the environment 250 of the vehicle 220. As shown in Figure 2, the vehicle 220 may be equipped with a variety of sensor systems 280. It should be noted that different sensor systems among the plurality of sensor systems 280 may be used to collect different types of data regarding the environment 250 of the vehicle 220.

[00069] In one example, the plurality of sensor systems 280 may include various optical systems, including, but not limited to , one or more camera-type sensor systems that are mounted on the vehicle 220 and communicatively coupled to the processor 102 of the electronic device 210. In general In other words, one or more camera-type sensor systems may be configured to collect image data from various parts of the environment 250 of the vehicle 220. In some cases, the image data provided by the one or more camera-type sensor systems may be used by the electronic device 210 to perform object detection procedures . For example, electronic device 210 may be configured to feed image data provided by one or more camera-type sensor systems to an object detection neural network (ODNN) that is trained to locate and classify potential objects in the environment 250 of vehicle 220.

[00070] In another example, the plurality of sensor systems 280 may include one or more radar-type sensor systems that are mounted on the vehicle 220 and communicatively coupled to processors 102. Generally speaking, one or more radar-type sensor systems may be configured to using radio waves to collect data about various parts of the environment 250 of the vehicle 220. For example, one or more radar-type sensor systems may be configured to collect radar data about potential objects in the environment 250 of the vehicle 220, such data potentially representing the distance of objects from the system radar type sensors, orientation of objects, velocity vector and/or speed of objects and the like.

[00071] In a further example, the plurality of sensor systems 280 may include one or more LIDAR systems that are installed on the vehicle 220 and communicatively coupled to the processors 102. Generally speaking, the LIDAR system is configured to capture data about the environment 250 of the vehicle 220. used, for example, to construct a multidimensional map of objects in the environment 250 of the vehicle 220. The LIDAR system may be installed on or retrofitted to the vehicle 220 in various locations and/or in various configurations to collect information about the environment 250 of the vehicle 220.

[00072] For example, depending on the implementation of the vehicle 220 and the LIDAR system, the LIDAR system may be mounted on the inside, top of the windshield of the vehicle 220. However, other locations for mounting the LIDAR system are within the scope of the present disclosure, including including on the rear window, side windows, front hood, roof, front grille, front bumper or side of the vehicle 220.

[00073] In the context of the present technology, the electronic device 210 is configured to detect one or more objects in the environment 250 of the vehicle 220 based on data received from one or more camera systems and from one or more LIDAR systems. For example, an electronic device 210 configured to detect a particular object in the environment 250 of a vehicle 220 may be configured to identify LIDAR data and camera data associated with that object, generate a "vector representation" representing features associated with that object, detecting an object by generating a bounding box for that object.

Training a machine learning algorithm

[00074] As noted above, machine learning algorithms must be trained before use. For many machine learning algorithms, this training is performed using a large set of pre-labeled training data. Preparing such a set of pre-labeled training data can be time-consuming and expensive, especially if the labels for each element in the data set are created by humans. Thus, in accordance with various implementations of the present technology, a method for automatically generating a training data set for a machine learning algorithm for use with an autonomous vehicle is described. This training data set can be generated using pre-collected data from a training vehicle that can be driven by a human driver. By analyzing the data collected by the learning vehicle, labels can be automatically generated and added to the data collected by the learning vehicle to generate a training data set.

[00075] FIG. 3 shows a timeline of the data collection process through a training vehicle (not shown) that can be controlled by a human operator (not shown) or that can be partially or fully automated. At each point in time (referred to as a "tick" of the process), the training vehicle receives an operational data element that includes data indicating the training location, the state of the training vehicle at the training location, and the state of the training object at the training location. The operational data item is then transferred to a store, which may include a database. Thus, as shown in timeline 300, during the operation of the training vehicle, on each tick from tick1 302 to tick 4 308, data is collected for work data item 322-328, which is transferred to storage 342-348. Although timeline 300 shows only four ticks 302-308, it should be understood that this data collection process may continue throughout the entire period of operation of the training vehicle, with an additional element of operational data being collected and stored at each tick during that period.

[00076] In some implementations, ticks such as ticks 302-308 may occur at predetermined regular time intervals. For example, the intervals between ticks could be 100ms, so there would be 10 ticks per second. However, ticks do not always have to be separated by equal periods of time. For example, in some implementations, the time spacing between ticks may depend on a vehicle parameter, such as vehicle speed. Thus, when the vehicle is moving at high speed, there may be less time between ticks than when the vehicle is moving at low speed.

[00077] In some implementations, each of the operational data elements 322-328 may include data about the training terrain, the state of the training vehicle, and the state of the training object. In some implementations, this data may be divided into multiple data sets, such as a training vehicle data set, a training object data set, and a training terrain data set.

[00078] The training vehicle state data may include various information, including, for example, the position of the training vehicle at the time of the tick, as well as other information about the condition of the training vehicle, such as training vehicle speed, direction, or angle turn, accelerator status, braking status, and other similar information.

[00079] In some implementations, the training locality may be part of the work data element for each tick, and in some implementations, data about the same training locality may be used for many ticks. In some implementations, the training terrain data may be terrain data read by sensors on the training vehicle. In some implementations, the training location data may be map data or data representing a bird's eye view of the training location.

[00080] In some implementations, the state of the training object may include the position of the training object during a tick and may include other data such as the speed and direction of the training object. In some implementations, the state of the training object may be represented by data collected by sensors on the training vehicle. It should be understood that the training object is typically a moving object of concern to the operation of the autonomous vehicle, such as a pedestrian, cyclist, or other vehicle. Additionally, it should be understood that while the examples show only one training object, in some implementations there may be multiple training objects at a training location in close proximity to the training vehicle.

[00081] In at least some embodiments of the present technology, operational data elements may include trajectory data associated with the training vehicle. Generally speaking, the trajectory data represents the trajectory that the training vehicle follows on the training road section. It is contemplated that the trajectory data may include multiple states of the training vehicle during operation that are associated with corresponding timestamps.

[00082] In some implementations, storage 342-348 may be storage in a single storage device or in a single database. In some implementations, each store 342-348 may be a separate record in a single database or multiple databases. It should be understood that storage 342-348 may be on a storage device or database local to the training vehicle, or may be external to the training vehicle, such as network or cloud storage. Additionally, it should be understood that all or part of the operational data may be transferred from storage 342-348 to a server or other system for use in training a machine learning algorithm.

[00083] FIG. 4 shows an exemplary data collection scenario by training vehicle 410 at training location 402. The training terrain 402 shown in the example of FIG. 4 includes a section of road forming a four-way intersection or road intersection. In the example shown in FIG. 4, training vehicle 410 and training object 412 (another vehicle in this example) arrive at the intersection at similar times. The training vehicle 410 crosses the intersection first.

[00084] In the example shown in FIG. 4, the training vehicle 410 and the training object 412 are shown as rectangles that represent the bounding boxes of the training vehicle 410 and the training object 412. In some implementations, the positions of the training vehicle 410 and the training object 412, as well as the distance between them, can be determined with using bounding boxes.

[00085] Line 420 shows the trajectory of the training vehicle, and line 422 shows the trajectory of the training object. Points 430, 432, 434, 436 and 438 show the positions of the training vehicle 410, respectively, at times t _-3 , t _-2 , t _-1 , t ₀ and t ₁ . Each of them represents the position of the vehicle during the corresponding "tick", as discussed above. Similar points 440, 442, 444, 446 and 448 show the positions of the training object 412, respectively, at times t _-3 , t _-2 , t _-1 , t ₀ and t ₁ . As can be seen, the paths of the training vehicle 410 and the training object 412 intersect at point 404.

[00086] The purpose of the training data that is collected by the training vehicle 410 is to train a machine learning algorithm to determine the intent parameter of a training object, such as training object 412, at a target time. In the example shown in FIG. 4, the target time point is time _t0 , so the state at time _t0 and any time before time _t0 can be used to form the training data set, while the state at time _t1 and any future times can be used to label the training data set with the appropriate intent parameter for what actually happened in the training scenario.

[00087] In some implementations, the states of the training vehicle 410 prior to the target time may be used to determine the trajectory or portion of the trajectory of the training vehicle 410 during a time interval of a predetermined duration prior to the target time. Likewise, in some implementations, the states of the training object 410 prior to the target time may be used to determine the trajectory or portion of the trajectory of the training object 410 during a time interval of a predetermined duration prior to the target time.

[00088] In some implementations, the training vehicle simply logs data for each tick. Once the data is logged, the data can be analyzed on the server (eg, the logged data can be transferred to the server after the training vehicle's training operation is completed) to determine the appropriate target times for use in the training data set. This can be done, for example, by finding states of the training vehicle and training object in the recorded data in which the training vehicle and training object occupied the same position or were within a predetermined threshold safe distance from each other in close proximity in time, e.g. within one tick or a small number of ticks. The target time point can then be found by going back to a previous point in time before the training vehicle occupied that position. It should be noted that while this can be done on the server at any time after the data is logged, in some implementations the learning vehicle may mark certain operational data items as having a high probability of being near such target times as the data is logged.

[00089] Thus, as seen in FIG. 4, both the training vehicle 410 and the training object 412 occupied a point 404 on the training terrain 402 during the interval between t ₀ and t ₁ . Returning to the nearest previously available tick identifies _t0 as the target time for training. Once this is established, the server can use the data from t ₀ and zero or more ticks before t ₀ to generate training sets, and the data from t ₁ and zero or more ticks after t ₁ to generate an accessory-free label representing the intent parameter regarding the training object. In some implementations, the training data may be divided into a training vehicle data set, a training object data set, and a training terrain data set.

[00090] Once the target time has been identified, the server may also generate an accessory-free label for the training data set based on one or more future states - i.e. states that occur after the target point in time - the learning vehicle and the learning object. An accessory-free label represents the intent of the training object at the target time. In other words, one or more future states are used as a marker indicating the intention at the target time.

[00091] In some implementations, an accessory-free label may be selected from a set of predefined future states for the training object. For example, in some implementations, the predefined future states for a learning object might be:

the training vehicle and the training object occupy one position on the training terrain after the target time, and that the training vehicle occupies this one position before the training object;

the training vehicle and the training object occupy one position on the training terrain after the target time, and that the training object occupies this one position before the training vehicle;

the training vehicle and the training object do not occupy the same position on the training terrain after the target time (for example, examples of “no interaction” for training purposes); And

the training object occupies a certain position on the training terrain at a future time, and that the training vehicle occupies another specific position on the training terrain at a future time, and the distance between said specific position and the other specific position is below a threshold safe distance.

[00092] Thus, in the example shown in FIG. 4, the accessory-free label that the system will generate based on the states of the training vehicle 410 and the training object 412 on the training terrain 402 at time t ₁ (and possibly at subsequent times) will indicate that the training vehicle and training object after the target time, occupy one position on the training terrain, and that the training vehicle occupies this one position before the training object. In some implementations, this accessorless label may be represented by a vector, with each of the predefined future states represented by a single value in the vector. For example, an accessory-free label for the example shown in FIG. 4 can be represented by a vector (1, 0, 0, 0).

[00093] It should be understood that although the scenario shown in FIG. 4 is based on the intersection of the training vehicle and the training object at the intersection, this is for illustration purposes only. Training a machine learning algorithm to predict the intent parameter of an object in the field will require many other scenarios and training datasets with non-accessorized labels. For example, there may be a variety of training scenarios in which a training vehicle and a training object are driving adjacent to each other in different lanes on the same road, with the training object and/or training vehicle performing a lane change. Many other training scenarios can also be used.

[00094] Referring now to FIG. 5, which describes a method for training a machine learning algorithm 502. This training typically occurs on the server based on stored operational data from many training runs of the training vehicle. As shown in FIG. 5, training data element 504 is sent to a machine learning algorithm 502, which generates a prediction 506 of the training object's intent parameter. The prediction 506 is then compared to an accessory-free label 508 that is associated with a training data item 504 (the generation of which is described above), and the difference between the prediction 506 and the accessory-free label 508 is then used to train a machine learning algorithm 502 .

[00095] In some implementations, training data element 504 may include a training vehicle (VV) data set 510 representing the state of the training vehicle at a target time point and at zero or more times before said target time point, data set 512 about a training object representing the state of the training object at a target time and at zero or more times before said target time, and a training terrain data set 514 representing the training terrain at a target time and at zero or more times before said target moment in time. In some implementations, the total training data set may include tens of thousands, hundreds of thousands, or even millions of such training items and their associated non-accessorized labels. In some implementations of the present technology, training data element 504 may include a training vehicle data set 510 representing the states of the training vehicle at a target time and 8 seconds prior to that target time.

[00096] As discussed above, in some implementations, an accessory-free label may be represented by a vector, with each value in the vector representing one of a predetermined number of states of the training vehicle and training object at a time point after the target time point. This vector represents the intent parameter of the training object. Similarly, in some implementations, prediction 506 may be represented by a similar vector, with each value in that vector representing a predicted probability (typically zero to one) for each of a predetermined number of training vehicle and training object states at a time point after the target time point (which may be referred to in this document as “intent classes”).

[00097] The machine learning algorithm 502 may be any suitable traditional machine learning algorithm. For example, the machine learning algorithm may be any of various traditional machine learning algorithms, including, but not limited to, "deep learning" algorithms, other types of neural networks or "connection" systems, decision trees, decision forests, Bayesian networks, or others known or more recently developed machine learning algorithms that use training data sets (for example, supervised or semi-supervised learning algorithms). The manner in which the prediction 506, the accessless label 508, and/or the difference between them is used to train the machine learning algorithm 502 will vary according to the machine learning algorithm being used.

[00098] After training the machine learning algorithm 502, information representing the machine learning algorithm may be sent to an electronic device associated with the autonomous vehicle, such as the electronic device 210 described above with reference to FIG. 2. The machine learning algorithm can then be used to control the self-driving vehicle.

Using a trained machine learning algorithm

[00099] FIG. 6 shows an example scenario for a machine learning algorithm used in a self-driving vehicle. As in FIG. 4, area 602 includes a four-way intersection or road intersection. An autonomous vehicle 610, including a trained machine learning algorithm, and an object 612 approach an intersection. The autonomous vehicle 610 uses sensors (not shown) to determine its own state and the detected state of the object 612 at each tick and stores at least enough previous data to apply a machine learning algorithm to predict the object's intent parameter on the ground. For example, if a machine learning algorithm has been trained to use data from a target time point and two previous ticks, then the autonomous vehicle 610 will retain state information from at least the most recent three ticks, so the machine learning algorithm can be applied. In the example shown in FIG. 4, the unmanned vehicle 610 has data about the state of the unmanned vehicle 610 and the state of the object 612 for at least times t ₀ , t _-1 , t _-2 , and t _-3 . The sensors also provide terrain information 602. In some implementations, this terrain data is recorded on every tick. In some implementations, terrain data may be recorded less frequently than state data. In some implementations, the terrain data may include predefined map data in addition to data provided by sensors on the autonomous vehicle 610.

[000100] The state data of the autonomous vehicle 610 may include information, such as the position of the autonomous vehicle 610 during a tick, as well as other information about the state of the autonomous vehicle 610, such as speed, direction or angle of rotation, accelerator state, braking state, and other information. The state data of the object 612 may include sensor information indicating a position, as well as other sensed information about the state of the object 612, such as the speed and direction of the object 612. As discussed above with reference to FIG. 4, the current and past states of the autonomous vehicle 610 and object 612 may be used in some implementations to determine their trajectories, or portions of such trajectories, corresponding to a time interval of a predetermined duration prior to a target time.

[000101] In FIG. 6, points 620, 622, 624 and 626 show the position of the unmanned vehicle 610, respectively, at times _t0 , t _-1 , t _-2 and t _-3 . Information about the state of the autonomous vehicle 610 at each of these times, and possibly previous times, forms an in-use data set 640 for the autonomous vehicle. Points 630, 632, 634 and 636 show the position of object 612, respectively, at times _t0 , t _-1 , t _-2 and t _-3 . Information about the state of the object 612 at each of these times, and possibly at previous times, forms the in-use data set 642 for the object. A set of in-use data (not shown) for terrain 602 may also be generated in the autonomous vehicle 610.

[000102] These data sets, during use, can be used by a trained machine learning algorithm to generate a predictive intent parameter for an object 612 in terrain 602. The autonomous vehicle 610 can then take an action based on this predictive intent parameter.

[000103] FIG. 7 depicts the generation of a predictive intent parameter for an object by a trained machine learning algorithm 702. The data element 704 at time of use is sent to a trained machine learning algorithm 702, which generates a predictive intent parameter 706 for the object.

[000104] The in-use data element 704 includes the most recent data about the state of the autonomous vehicle, about the state of the object, and about the terrain. In some implementations, in-use data element 704 may include in-use data set 710 for an unmanned vehicle (UV), in-use data set 712 for an object, and in-use data set 714 for a terrain. As discussed above, each of these data sets during use may include data for the target time point as well as zero or more previous time points depending on the trained machine learning algorithm.

[000105] In some implementations, the predictive intent parameter 706 may be represented by a vector, with each value in the vector representing a predicted probability (typically zero to one) for each of a predetermined number of autonomous vehicle and object states at a time point after the target time point (which may be referred to in this document as “intent classes”). This vector provides a prediction of the object's intent in the near future and can be used to determine actions that should be taken by the autonomous vehicle, for example, to avoid a collision or other unwanted event.

[000106] The trained machine learning algorithm 702 may be any suitable traditional machine learning algorithm that has been trained as described above. For example, the machine learning algorithm may be any of various traditional machine learning algorithms, including, but not limited to, "deep learning" algorithms, other types of neural networks or "connection" systems, decision trees, decision forests, Bayesian networks, or others known or later developed machine learning algorithms.

Overview of how to train and use a machine learning algorithm

[000107] Referring now to FIG. 8, which describes a flowchart 800 of a method for training a machine learning algorithm for predicting a terrain object's intent parameter in accordance with the present technology. The method is divided into two phases: a training phase 850 and a use phase 852. The training phase 850 is typically performed by the server, while the use phase is performed on an electronic device associated with the autonomous vehicle, such as electronic device 210, as described with reference to FIG. 2.

[000108] At block 802 of training phase 850, the server accesses operating data associated with the training vehicle, wherein the operating data includes data indicating the training location, the state of the training vehicle at the training location, and the state of the training object at the training location. The operating data may have been previously stored in a storage device such as a storage device or a database.

[000109] At block 804 of training phase 850, the server retrieves data for the target time point from the operational data. The extracted data indicates: the training location at the target time; a target state of the training vehicle corresponding to the target time point, and a future state of the training vehicle after the target time point; and a target state of the learning object corresponding to the target time point, and a future state of the learning object after the target time point.

[000110] In some implementations, the server selects a target time point in the data by identifying a certain state of the training vehicle and a certain state of the training object in the operating data that are associated with one position on the training location. The server may then select the target time as a prior time relative to a specific time associated with a specific state of the training vehicle.

[000111] In some implementations, retrieving data for a target time point from the operational data includes retrieving data indicating a set of previous states of the training vehicle prior to the target time point. Data retrieval may also include retrieving a set of previous states of the training object prior to the target time. In some implementations, the set of previous states of the training vehicle is a portion of the training vehicle's trajectory to a target time, wherein the portion of the trajectory corresponds to a time interval of a predetermined duration to the target time. In some implementations, the set of previous states of the training object is part of the trajectory of the training object to the target time, and part of the trajectory corresponds to a time interval of a predetermined duration to the target time.

[000112] In some implementations, data retrieval includes retrieving data indicating a set of future states of the training vehicle after a target time, including said future state of the training vehicle, and a set of future states of the training object after the target time, including the mentioned future state of the learning object. In some implementations, the set of future states of the training vehicle is part of the trajectory of the training vehicle after the target time, and the set of future states of the training object is part of the trajectory of the training object after the target time.

[000113] At block 806 of training phase 850, the server generates a training set for a target time point for training the machine learning algorithm, wherein the training set has input information and an accessory-free label. The input information includes data indicating the training location and the state of the training vehicle and training object at the target time. An accessory-free label is based on data indicating the future states of the learning vehicle and learning object. An accessory-free label indicates the intent of the training object at the training location at the target time.

[000114] In some implementations, the server generates an accessory-free label based on data indicating future states of the training vehicle and training object. In some implementations, generating an accessoryless label may include determining that the future states of the training vehicle and the training object indicate at least one of: the training vehicle and the training object, after the target time, occupy the same position on the training terrain, and that the training vehicle takes this one position before the training object; the training vehicle and the training object occupy one position on the training terrain after the target time and that the training object occupies this one position before the training vehicle; the training vehicle and the training object do not occupy the same position on the training area after the target time; and that the training vehicle occupies a certain position on the training terrain at a future time point, and that the training vehicle occupies another specific position on the training terrain at a future time point, and the distance between said specific position and the other specific position is below a threshold safe distance.

[000115] In at least some embodiments, it should be noted that at least some of the training examples generated by the server for training the machine learning algorithm may contain examples of "no interaction" between the training vehicle and the training object. In some cases, if the training examples only have examples of "interaction" between the training vehicle and the training object, the machine learning algorithm can predict intent parameters that always indicate some interaction between the vehicle during use and the object during use, which is maybe not so.

[000116] In block 808 of training phase 850, the server trains a machine learning algorithm based on the training set to predict, during the use phase, an intent parameter indicating the intent of the terrain object at the time of the target use time. The machine learning algorithm is trained to base this prediction on data indicating the terrain at the target use time, the during use state of the autonomous vehicle at the target use time, and the during use state of the object at the target use time.

[000117] Blocks 802-808 may be repeated until the machine learning algorithm's prediction accuracy reaches a predetermined sufficient level of accuracy. Once this level of prediction accuracy is achieved, in training phase block 810, the server transmits information representing the trained machine learning algorithm to an electronic device associated with the autonomous vehicle.

[000118] At block 812 of use phase 852, the electronic device associated with the autonomous vehicle executes the trained machine learning algorithm to generate an object intent parameter. Generation of the intent parameter is based on data indicating the terrain at the target use time, the during use state of the autonomous vehicle at the target use time, and the during use state of the object at the target use time. In some implementations, the intent parameter indicates the likelihood that the entity's intent is one of a plurality of intent classes, and the machine learning algorithm is configured to classify the entity's intent into the plurality of intent classes.

[000119] At block 814 of use phase 852, the electronic device initiates operation of the autonomous vehicle based at least in part on the intent parameter.

[000120] Modifications and improvements to the above-described implementations of the present technology may become apparent to those skilled in the art. The foregoing description is intended to be exemplary and not limiting. It is therefore intended that the scope of the present technology be limited only by the scope of the appended claims.

Claims (76)

1. A method for training a machine learning algorithm to predict an intent parameter for an object on the ground for use in unmanned vehicles, wherein the unmanned vehicle moves through the terrain in the vicinity of said object, the method being executed by a server and comprising: allowing the server to access operational data associated with the training vehicle, wherein the operational data has previously been stored in storage and includes data indicating: the training area, the condition of the training vehicle at the training area and the condition of the training object at the training area; for the target time in the operating data: retrieving from operational data by means of a server data indicating: training area at the target time, the target state of the training vehicle corresponding to the target time point, and the future state of the training vehicle after the target time point, the target state of the learning object corresponding to the target time point, and the future state of the learning object after the target time point; generation by the server of a training set for training a machine learning algorithm, wherein the training set has input information and an accessory-free label, wherein the input information includes data indicating the training location and target states of the training vehicle and training object, the non-assessed label is based on data indicating future states of the training vehicle and the training object, the non-assessed label indicating the intention of the training object at the training location at the target time; And server training of a machine learning algorithm based on the training set for prediction, during the use phase: an intent parameter indicating the intent of an object on the terrain at the target use time, based on data indicating (i) the terrain at the target use time, (ii) the use state of the unmanned vehicle at the target use time, and (iii) the use state of said object. 2. The method according to claim 1, in which the extract contains: retrieving by the server data indicating (i) a set of previous states of the training vehicle up to the target time point and (ii) a set of previous states of the training vehicle up to the target time point, wherein the input information further includes data indicating sets of previous states of the training vehicle and the training object. 3. The method according to claim 2, wherein the set of previous states of the training vehicle is a part of the trajectory of the training vehicle up to the target time, wherein the part of the trajectory corresponds to a time interval of a predetermined duration up to the target time. 4. The method according to claim 2, in which the set of previous states of the learning object is part of the trajectory of the learning object up to the target point in time, and part of the trajectory corresponds to a time interval of a predetermined duration up to the target point in time. 5. The method according to claim 1, in which the extract contains: retrieving by the server data indicating (i) a set of future states of the training vehicle after the target time, including said future state of the training vehicle, and (ii) a set of future states of the training object after the target time, including said future state of the training object, The accessoryless label is based on the sets of future states of the learning vehicle and the learning object. 6. The method of claim 5, wherein the set of future states of the training vehicle is part of a trajectory of the training vehicle after the target time, and wherein the set of future states of the training object is part of another trajectory of the training object after the target time. 7. The method of claim 1, wherein the method further comprises selecting by the server a target time point in the operational data, the selection comprising: identifying by the server a certain state of the training vehicle and a certain state of the training object in the operating data that are associated with the same position on the training area; And the server selecting a target time point as a prior time point relative to a certain time point associated with a certain state of the training vehicle. 8. The method of claim 1, wherein the method comprises generating, by the server, an accessory-free label based on data indicating future states of the training vehicle and the training object, wherein generating the accessory-free label includes: determining by the server that the future states of the training vehicle and training object indicate at least one of the following: the training vehicle and the training object occupy one position on the training terrain after the target time, and that the training vehicle occupies this one position before the training object; the training vehicle and the training object occupy one position on the training terrain after the target time and that the training object occupies this one position before the training vehicle; the training vehicle and the training object do not occupy the same position on the training area after the target time; And the training object occupies a certain position on the training terrain at a future time, and that the training vehicle occupies another specific position on the training terrain at a future time, and the distance between said specific position and the other specific position is below a threshold safe distance. 9. The method according to claim 1, wherein the method additionally contains: transmitting by the server information representing the machine learning algorithm to an electronic device associated with the unmanned vehicle; at the target time of use, during operation of the unmanned vehicle: executing a machine learning algorithm by the electronic device to generate an intent parameter; And causing the electronic device to initiate operation of the autonomous vehicle based at least in part on an intent parameter. 10. The method of claim 1, wherein the machine learning algorithm is configured to classify the intent of an object, wherein the intent parameter indicates the likelihood that the intent of the object is one of a plurality of intent classes. 11. The method according to claim 1, wherein the training vehicle is an unmanned vehicle. 12. The method according to claim 1, wherein the object is a moving object. 13. The method of claim 1, wherein the object is at least one of a vehicle and a pedestrian. 14. The method of claim 1, wherein the training location comprises a training road section, and wherein the training location data is a bird's eye view image of the training road section at a target time. 15. A method of operating an unmanned vehicle, containing: receiving, at an electronic device associated with the unmanned vehicle, data from a sensor associated with the unmanned vehicle, the data indicating the state of the unmanned vehicle, the state of an object in the vicinity of the unmanned vehicle, and the terrain over which the unmanned vehicle is moving; providing said data to a machine learning algorithm on the electronic device, wherein the machine learning algorithm determines an intent parameter indicating the intent of the object on the ground at the time of the target time; And causing the electronic device to operate the autonomous vehicle based at least in part on the intent parameter; wherein the machine learning algorithm was trained using a training data set generated by, at least in part: the server accessing operational data associated with the training vehicle, wherein the operational data has previously been stored in storage and includes data indicating: the training area, the condition of the training vehicle at the training area and the condition of the training object at the training area; for the target training time in the operating data: extracting from operational data by means of a server data indicating: training area at the target time of training, a target state of the training vehicle corresponding to the target training time point, and a future state of the training vehicle after the target training time point, the target state of the training object corresponding to the target moment in time of training, and the future state of the training object after the target moment in time of training; generating by the server a training data set for training the machine learning algorithm, wherein the training data set has input information and an accessory-free label, wherein the input information includes data indicating the training location and target states of the training vehicle and training object, the non-assessed label is based on data indicating the future states of the training vehicle and the training object, the non-assessed label indicating the intention of the training object at the training location at the training target time. 16. The method of claim 15, wherein the machine learning algorithm is configured to classify the intent of an object, wherein the intent parameter indicates the probability that the intent of the object is one of a plurality of intent classes. 17. The method according to claim 15, further comprising receiving on the electronic device information representing the machine learning algorithm from the server. 18. An unmanned vehicle containing: at least one sensor configured to collect sensor data indicating the state of the unmanned vehicle, the state of an object in the vicinity of the unmanned vehicle, and the terrain over which the unmanned vehicle is moving; an electronic device including a processor and a memory, the memory storing programmed instructions that, when executed by the processor, cause the electronic device to: provide the sensor data to a machine learning algorithm running on the processor, wherein the machine learning algorithm determines an intent parameter indicating the intent of an object on the ground at the time of the target time; And initiate operation of the autonomous vehicle based at least in part on the intent parameter; wherein the machine learning algorithm contains values stored in memory that were determined by training the machine learning algorithm using a training data set generated by, at least in part: the server accessing operational data associated with the training vehicle, wherein the operational data has previously been stored in storage and includes data indicating: the training area, the condition of the training vehicle at the training area and the condition of the training object at the training area; for the target training time in the operating data: extracting from operational data by means of a server data indicating: training area at the target time of training, a target state of the training vehicle corresponding to the target training time point, and a future state of the training vehicle after the target training time point, the target state of the training object corresponding to the target moment in time of training, and the future state of the training object after the target moment in time of training; generating by the server a training data set for training the machine learning algorithm, wherein the training data set has input information and an accessory-free label, wherein the input information includes data indicating the training location and target states of the training vehicle and training object, the non-assessed label is based on data indicating the future states of the training vehicle and the training object, the non-assessed label indicating the intention of the training object at the training location at the training target time. 19. The autonomous vehicle of claim 18, wherein the machine learning algorithm is configured to classify the intent of an object, wherein the intent parameter indicates the probability that the object's intent is one of a plurality of intent classes. 20. The self-driving vehicle of claim 18, wherein the memory further stores programmed instructions that, when executed by the processor, cause the electronic device to receive information representing the machine learning algorithm from the server.