CN110264586A - L3 grades of automated driving system driving path data acquisitions, analysis and method for uploading - Google Patents

Abstract

The invention relates to a method for collecting, analyzing and uploading road driving data of an L3-level automatic driving system. End-to-end driving data for online data analysis, including automatic driving system intermediate result output interface definition, target matching consistency detection, positioning road sign semantic output, extreme vehicle operation detection and human-machine decision consistency detection; Ready to upload; the server receives and stores the driving data of the vehicle. The invention meets the development and verification requirements of the L3-level automatic driving system perception, positioning and planning decision-making modules; the automatic driving extreme scene detection greatly reduces the bandwidth occupied by data recording and uploading; the corresponding algorithm modules can be run online to maximize the front-end platform data. Mining and verification greatly reduces the human resource requirements for post-processing data screening.

Description

L3-level automatic driving system road driving data collection, analysis and upload method

technical field

The invention relates to an automobile automatic driving system, in particular to a method for collecting, analyzing and uploading road driving data of an L3-level automatic driving system.

Background technique

Intelligence is one of the important trends in the development of the automotive industry today. It is expected that intelligent driving technologies and systems will develop rapidly around the world from 2020 to 2030. The automatic driving system is divided into six levels from low to high according to the degree of intelligence, of which the L3 level of automatic driving is defined as allowing the corresponding system to replace the driver to drive the vehicle independently in the defined driving scenario, such as reducing the speed in high-speed driving scenarios. Driving burden, etc. At present, the L1-level and L2-level advanced assisted driving systems have been implemented in some mass-produced models, and the L3-level autonomous driving system is still in the prototype development stage, and a lot of testing and verification work is needed.

Compared with L1-level and L2-level assisted driving systems, the application scenarios of L3-level autonomous driving systems are more complex, and the amount of driving data required for development and verification is larger. Machine learning methods are more widely used in L3 and above systems, so the demand for valid data for corresponding driving scenarios has also doubled. Compared with the L1-level and L2-level assisted driving systems, the data transmission and calculation amount of the L3-level automatic driving system are doubled. Therefore, extracting the effective data required for L3-level autonomous driving systems in road driving scenarios has important practical application value for the industrialization of such systems. Real-time uploading of L3 autonomous driving system operating scene data is not feasible even in the 5G era. It not only requires a large amount of transmission bandwidth, but also requires a lot of subsequent manpower to classify and verify the above driving scenes.

Most of the existing on-board data recording systems are based on CAN bus data recording. Most of these systems are local recordings of the vehicle, and the required data bandwidth and storage space are very limited, so they have no reference significance for the data recording of automatic driving systems. Some rear-mounted driving data recording devices (for operation and commercial vehicles) can record multi-channel video stream data (inside and outside the vehicle), and correlate some CAN bus vehicle data (vehicle speed, etc.) and GPS information. However, such devices do not have or utilize front-end computing power, and the recorded data needs to undergo a lot of post-processing before they can be used for some vision system development and testing (offline development mode).

The existing vehicle driving record system data storage and transmission methods have the following shortcomings: (i) the development and testing data required for L3 and above autonomous driving systems cannot be fully recorded; (ii) a large amount of redundancy will be recorded, and system development and verification Data that does not help much consumes unnecessary data transmission and storage resources; (iii) It needs to consume a lot of human resources to screen and extract valid data offline, and can only verify the problem offline; (iv) After-installed equipment and actual vehicle computing platform There are differences in computing characteristics and capabilities, and algorithm modules cannot be iteratively tested online.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the present invention provides a method for collecting, analyzing and uploading road driving data of an L3-level automatic driving system, which can achieve the following purposes: (i) satisfy the requirements of each module of the L3-level automatic driving system perception, positioning and planning decision-making. Development and verification requirements; (ii) automatic driving extreme scene detection, greatly reducing the bandwidth occupied by data recording and uploading; (iii) online operation of corresponding algorithm modules, maximizing front-end platform data mining and verification, and greatly reducing post-processing data screening work Human resource needs.

The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions: the L3-level automatic driving system road driving data collection, analysis and upload method of the present invention comprises the following steps:

①Car-side driving data collection;

②On-line data analysis of the collected vehicle-end driving data;

③Data communication, prepare for uploading of vehicle-side driving data;

④The server side receives and stores the driving data of the vehicle side.

The present invention is based on on-board WIFI or 4G network communication and edge intelligent computing, and can meet the data recording requirements required for the development and verification of L3-level automatic driving functions. Vehicles equipped with L3-level autonomous driving systems are equipped with cameras or cameras (vision), millimeter-wave radars, integrated navigation equipment and in-vehicle data processing terminals. The in-vehicle data processing terminals include a data acquisition module, an intelligent analysis module, and a communication module. a component. The data that can be automatically collected and uploaded by the invention includes: positioning road sign data, accident scene data, vehicle extreme operation scene data, perception matching abnormal scene data, human-machine response mismatch scene data, and custom request data. The invention meets the development and verification requirements of the L3-level automatic driving system perception, positioning and planning decision-making modules; the automatic driving extreme scene detection greatly reduces the bandwidth occupied by data recording and uploading; the corresponding algorithm modules can be run online to maximize the front-end platform data. Mining and verification greatly reduces the human resource requirements for post-processing data screening.

Preferably, the step 1 includes the following steps:

(11) Data acquisition and synchronization: by software synchronization, each driving data is synchronized by collecting GPS clock or vehicle terminal system clock to construct driving data including time, image data, radar original data, integrated navigation data and vehicle dynamics parameters structure;

(12) Data encoding and caching: The image data is encoded by H264 or H265, and other data is encoded by virtual CAN information; data registers are set to cache the driving data structure.

Preferably, the step ② includes the following steps: (21) definition of an automatic driving system intermediate result output interface; (22) target matching consistency detection; (23) semantic output of localized road signs; (24) extreme vehicle operation detection; ( 25) Consistency detection of human-machine decision-making.

Preferably, the specific method for the target matching consistency detection in the step (22) is:

According to the output results of the millimeter-wave radar and the visual perception system, extract the driving data whose timing matching distance tolerance is greater than the preset threshold dmin, and calculate the matching distance according to the following formula:

in, output target coordinates for the radar, Output the target coordinates for the on-board camera, n is the target life cycle; if d>dmin, upload the driving data of the corresponding segment.

Preferably, the specific method for detecting extreme vehicle operation in the step (24) is:

The yaw rate, longitudinal acceleration, longitudinal deceleration and lateral acceleration output by the inertial navigation system are time-series encoded, and the time-series encoded data are classified into extreme vehicle operations by numerical analysis methods or machine learning methods. Extreme vehicle operations are classified into rapid acceleration operations. , sharp deceleration operation and sharp turning operation;

The numerical analysis method is as follows: if the measured value of longitudinal acceleration is continuously greater than the set threshold value A2min for more than N times, it is confirmed as a rapid acceleration operation; if the measured value of longitudinal deceleration is continuously less than the set threshold value A2min for more than N times, it is confirmed as a rapid deceleration operation ; If the measured value of lateral acceleration and the measured value of yaw rate are continuously greater than the set threshold AYmin and Tmin for more than N times, it is confirmed as a sharp turning operation;

The machine learning method is as follows: using extreme vehicle operation and driving time series data samples, offline training support vector machine or long short-term memory network; deploying the trained model above in the vehicle analysis terminal, the input is time-series encoded inertial navigation data, and the output is rapid acceleration operation. , an event signal for a sharp deceleration operation or a sharp turn operation.

Preferably, the concrete method of the man-machine decision consistency detection in the step (25) is:

According to the output result of the planning layer and the output result of the vehicle pose, according to the preset preview distance, extract the driving data segment whose deviation between the actual trajectory and the planned trajectory is greater than the preset threshold Dmin; calculate the normalization in the vehicle coordinate system according to the following formula Trajectory deviation:

Among them, [Xi, Yi] is the actual track point coordinates in the vehicle coordinate system, [xi, yi] is the planned track point coordinates in the vehicle coordinate system; m is the number of track points; if D>Dmin, upload the driving data of the corresponding segment.

Preferably, in the step (21), the specific method for defining the intermediate result output interface of the automatic driving system is: including sensing target output, positioning road sign semantic output, vehicle pose output and planning decision output;

Among them, the perception target output: including the target output of the vision system, the target output of the millimeter wave radar system, and the original target point cloud output of the millimeter wave radar system;

Semantic output of localized road signs: Output the semantics of localized road signs in the form of binary layers, including the semantic output of drivable areas, the semantic output of lane boundaries and the semantic output of indicative road signs;

Vehicle pose output: including vehicle position, vehicle speed, heading angle and 6-axis inertial sensor output;

Planning decision output: The preview point trajectory is given at fixed longitudinal space intervals.

Preferably, the specific method for locating the landmark semantic output in the step (23) is:

According to the location semantic output of the perception module, and use the vehicle pose estimation result of the location module to construct the location landmark semantic output; the location landmark semantic output method adopts the key frame semantic output method or the compressed full semantic output method;

The key frame semantic output method is: according to the mileage integral estimation of the positioning module, extract a frame of key semantic output every 50 meters, and construct a key frame semantic location register. The key frame semantic location register stores the key frame and the vehicle latitude and longitude data at the corresponding time; Packets are compressed every 20 frames, and an upload request signal is sent;

The compressed full-semantic output method is: including lane-level semantics and guidance-level semantics; post-processing to extract the number of lanes, lane widths, boundary types, lanes, and off-center distances from the visual system lane and drivable area semantic output layers; The post-processing of the road sign semantic output layer extracts two parts of the road sign and the space road sign; the above data and the latitude and longitude data of the vehicle at the corresponding time are packaged once every 1km traveled, and an upload request signal is sent.

Preferably, the specific method of data communication in step (3) is: according to the online data analysis result in step (2), compress the driving data queue obtained after collecting in step (1), and name the corresponding compressed file according to predefined rules; use TCP Or UDP protocol, transparently transmit compressed data to the server through 4G network or wireless network;

The specific method for receiving and storing the vehicle-side driving data in the step (4) is as follows: on the server side, through the TCP or UDP protocol, the driving data sent by the vehicle-mounted terminal is received; the data collection date is used as a subdirectory to store the relevant driving data.

The beneficial effects of the present invention are as follows: the present invention is based on on-board WIFI or 4G network communication and edge intelligent computing, filters data streams that are not significant for system optimization and upgrading in most driving scenarios, and can collect compressed road sign data required for automatic driving positioning , accident scene data, scene data under specific extreme vehicle operation, abnormal scene data matching the detection results of radar and camera, and scene data with large differences in human-machine response. The present invention can achieve the following effects: (i) meet the development and verification requirements of the L3 level automatic driving system perception, positioning, planning and decision-making modules; (ii) automatic extreme scene detection, greatly reducing the bandwidth occupied by data recording and uploading; (iii) ) can run the corresponding algorithm modules online, maximize the data mining and verification of the front-end platform, and greatly reduce the human resource requirements for post-processing data screening.

Description of drawings

FIG. 1 is a schematic top view of a vehicle in the present invention.

FIG. 2 is a flow chart of an algorithm of the present invention.

Detailed ways

The technical solutions of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

Embodiment: The method for collecting, analyzing, and uploading road driving data of an L3-level autonomous driving system in this embodiment is based on vehicle WIFI or 4G network communication and edge intelligent computing, and can meet the data required for the development and verification of L3-level autonomous driving functions Document requirements. A car equipped with an L3-level autonomous driving system is shown in Figure 1. The car is equipped with a camera or camera (vision), millimeter-wave radar, integrated navigation equipment, and on-board data processing terminal. The on-board data processing terminal includes three parts: data acquisition module, intelligent analysis module and communication module. Among them, the data acquisition module mainly includes various data interfaces, acquisition chips (MCU) and video encoding modules, which are responsible for collecting, synchronizing and encoding various driving data; the positioning module mainly includes GPS and inertial navigation modules, responsible for positioning road sign extraction and event recording. Association with the map; the intelligent analysis module mainly includes the L3 level automatic driving system processing terminal and the vehicle intelligent analysis terminal (integrated with multi-core arm and neural network acceleration unit), which is responsible for real-time processing of scene data streams, and selects the scene data to be recorded according to preset rules. The communication module includes 4G and WIFI modules, which are mainly responsible for compressing and uploading the corresponding encoded driving data.

The method for collecting, analyzing and uploading road driving data for L3 autonomous driving system, as shown in Figure 2, includes the following steps:

①Car-side driving data collection:

Vehicle-side driving data consists of scene data (raw data of radar and vision systems), position and attitude information (integrated navigation device data), and vehicle dynamics data (vehicle speed, accelerator, braking and steering input); vehicle-side driving data collection includes: The following steps:

(11) Data acquisition and synchronization: by software synchronization, each driving data is synchronized by collecting GPS clock or vehicle terminal system clock to construct driving data including time, image data, radar original data, integrated navigation data and vehicle dynamics parameters Structure; the driving data structure contains the following properties:

Time: the time point corresponding to the driving data;

Image data: 8 channels of original image input (Imgl-Img8);

Radar raw data: radar system output target list (the default maximum output number is 32: Obj1-Obj32);

Integrated navigation data: including latitude and longitude and vehicle 6-DOF motion information (three-axis acceleration and three-axis angular velocity);

Vehicle dynamics parameters: including vehicle speed, steering wheel angle, accelerator travel and braking travel, etc.;

(12) Data encoding and caching: The image data is encoded in H264 or H265, and other data is encoded in virtual CAN information (64-bit); set the data register, the length can be configured (the default is 300, that is, 12 seconds) 25fps data or 10 seconds of 30fps data) to cache the driving data structure.

②Online data analysis of the collected vehicle-end driving data:

The L3-level autonomous driving system processes the perception, fusion, positioning, and planning decision-making algorithm modules running on the terminal, and outputs the results of each custom intermediate level. After post-processing by the intelligent on-board analysis terminal, it sends data upload instructions to the communication module according to the upload type input by the user. ; details include the following steps:

(21) The definition of the intermediate result output interface of the automatic driving system, the specific methods are: including the output of the perception target, the semantic output of the positioning road sign, the output of the vehicle pose and the output of the planning decision;

Among them, perceptual target output: including vision system (front-sighted, blind-spot and rear-sighted, a single camera has a default upper limit of 16 targets, attributes include target type, vertical distance, horizontal distance and relative speed, etc.) target output, millimeter-wave radar system (front and rear 776Hz) Radar and blind spot 24GHz radar, single radar default target upper limit is 16, attributes include radial distance, angle and relative speed, etc.) target output and millimeter wave radar system original target point cloud output (single radar default point target number 150, attributes include reflectivity, radial distance, angle, and relative velocity, etc.);

Semantic output of localized road signs: Output the semantics of localized road signs in the form of binary layers, including the semantic output of drivable areas, the semantic output of lane boundaries and the semantic output of indicative road signs;

Vehicle pose output: including vehicle position (latitude and longitude or the plane position in the custom initialization world coordinate system), vehicle speed, heading angle and 6-axis inertial sensor output (lateral, longitudinal, lateral acceleration and yaw, pitch, roll angular velocity) ;

Planning decision output: The trajectory of preview points (default 10 preview points) is given according to a fixed vertical space interval (default 1 meter);

(22) Target matching consistency detection, the specific method is:

According to the output results of the millimeter-wave radar and the visual perception system, extract the driving data whose timing matching distance tolerance is greater than the preset threshold dmin, and calculate the matching distance according to the following formula:

in, output target coordinates for the radar, Output target coordinates for the on-board camera, n is the target life cycle; if d>dmin, upload the driving data of the corresponding segment;

(23) Locating landmark semantic output, the specific method is:

According to the location semantic output of the perception module, and use the vehicle pose estimation result of the location module to construct the location landmark semantic output; the location landmark semantic output method adopts the key frame semantic output method or the compressed full semantic output method;

The key frame semantic output method is: according to the mileage integral estimation of the positioning module, extract a frame of key semantic output (ie, the binary semantic output layer of the visual system after post-processing) every 50 meters, and construct the key frame semantic positioning register. The frame semantic location register stores the key frame and the latitude and longitude data of the vehicle at the corresponding time; the packet is compressed once every 20 frames (that is, every 1km of driving), and an upload request signal is sent;

The compressed full-semantic output method is: including lane-level semantics and guidance-level semantics; post-processing to extract the number of lanes, lane widths, boundary types, lanes, and off-center distances from the visual system lane and drivable area semantic output layers; Instruct the road sign semantic output layer to post-process to extract two parts of data of road signs and spatial road signs; the above data together with the latitude and longitude data of the vehicle at the corresponding time are compressed once every 1km of driving, and an upload request signal is sent;

(24) Extreme vehicle operation detection, the specific method is:

Encode the yaw rate, longitudinal acceleration, longitudinal deceleration and lateral acceleration output by the inertial navigation system in time series, and use numerical analysis method or machine learning method (SVM or LSTM) to classify the time series encoded data for extreme vehicle operation, extreme vehicle operation Divided into rapid acceleration operation, rapid deceleration operation and sharp turning operation;

The numerical analysis method is as follows: if the measured value of longitudinal acceleration is continuously greater than the set threshold value A1min for more than N times (the default value of N is 3), it is confirmed as a rapid acceleration operation; if the measured value of longitudinal deceleration is continuously less than the set threshold value A2min for more than N times (The default value of N is 3), it is confirmed as a sudden deceleration operation; if the measured value of lateral acceleration and the measured value of yaw rate are continuously greater than the set threshold AYmin and Tmin for more than N times (the default value of N is 3), it is confirmed as sharp turn operation;

The machine learning method is: using extreme vehicle operation and driving time series data samples, offline training support vector machine (SVM) or long short-term memory network (LSTM); deploy the trained model above in the vehicle analysis terminal (acceleration, deceleration binary classification and Two classifications of sharp turns), the input is the time-series encoded inertial navigation data, and the output is the event signal of the sudden acceleration operation, the sudden deceleration operation or the sharp turning operation;

(25) Human-machine decision-making consistency detection, the specific method is:

According to the output result of the planning layer and the output result of the vehicle pose, according to the preset preview distance, extract the driving data segment whose deviation between the actual trajectory and the planned trajectory is greater than the preset threshold Dmin; calculate the normalization in the vehicle coordinate system according to the following formula Trajectory deviation:

Among them, [Xi, Yi] is the actual track point coordinates in the vehicle coordinate system, [xi, yi] is the planned track point coordinates in the vehicle coordinate system; m is the number of track points, the default is 10; if D>Dmin, upload the corresponding segment driving data;

(26) Custom data request: according to the request signal input of the human-computer interaction port, send a data upload instruction according to a preset rule, that is, upload the driving data of the register in step (12) at the request time;

③Data communication, prepare for uploading of the driving data on the vehicle side, the specific methods are as follows:

According to the online data analysis result in step ②, compress the driving data queue in the register in step (12) (Lz4 can be used for data compression), and name the corresponding compressed file according to the predefined rules; establish the on-board data processing terminal as a server , using TCP or UDP protocol to transparently transmit compressed data to the server through 4G network or wireless network;

④The server side receives and stores the driving data of the vehicle side, and the specific method is as follows:

On the storage server (cloud), a client is established to receive the driving data sent by the vehicle data processing terminal through TCP or UDP protocol; the relevant driving data is stored in the NTFS file system with the data collection date as a subdirectory.

The present invention is based on on-board WIFI or 4G network communication and edge intelligent computing, and can automatically collect and upload the following driving data in the road driving scenario of a human-driven vehicle:

1. Positioning road sign data: As an option, extract structured semantics (including road road signs and space road signs, etc.) for the positioning road signs that define the L3 system function application scenarios (including parking scenarios and high-speed scenarios, etc.), according to the predefined data. The structure is compressed and uploaded to the server (cloud).

2. Accident scene data: as an option, identify the vehicle collision state according to the collision sensor signal. Save the corresponding scene data according to the predefined accident recording rules.

3. Extreme vehicle operation scene data: as an option, identify extreme vehicle dynamics states, such as sharp turns, sharp acceleration and deceleration, based on 3-axis or 6-axis inertial navigation system (gyroscope) measurement data. Record and upload the corresponding scene data according to the predefined event association rules.

4. Perceived matching abnormal scene data: as an option, according to the scene perception matching results of the millimeter-wave radar and the vision system, according to the preset target matching tolerance (the distance of the overlooking vehicle coordinate system or the coincidence of the image coordinate system target), record and upload the corresponding scene data.

5. The human-machine response does not match the scene data: as an option, run the "virtual driver" on the in-vehicle computing platform, that is, the local trajectory planning algorithm module, to match the real vehicle kinematics state (that is, the real driver's vehicle operation), press The preset rules record the driving scene data that the similarity of human-machine trajectory does not meet the requirements.

6. Customized request data: As an option, the driver/tester input interface is provided on the interactive terminal, and the current driving scene data can be requested to record the current driving scene data with one click in a preset method (the default is the full record of the driving data of the preset duration).

Based on on-board WIFI or 4G network communication and edge intelligent computing, the invention filters out data streams that are not significant for system optimization and upgrade in most driving scenarios, and can collect compressed road sign data, accident scene data, and specific limit data required for automatic driving positioning. The scene data under vehicle operation, radar and camera detection results match abnormal scene data and scene data with large differences in human-machine responses. The present invention can achieve the following effects: (i) meet the development and verification requirements of the L3 level automatic driving system perception, positioning, planning and decision-making modules; (ii) automatic extreme scene detection, greatly reducing the bandwidth occupied by data recording and uploading; (iii) ) can run the corresponding algorithm modules online, maximize the data mining and verification of the front-end platform, and greatly reduce the human resource requirements for post-processing data screening.

Claims (9)

1. a L3 level automatic driving system road driving data collection, analysis and uploading method, is characterized in that comprising the following steps: ①Car-side driving data collection; ②On-line data analysis of the collected vehicle-end driving data; ③Data communication, prepare for uploading of vehicle-side driving data; ④The server side receives and stores the driving data of the vehicle side. 2. the L3 level automatic driving system road driving data collection, analysis and uploading method according to claim 1, is characterized in that described step 1. comprises the following steps: (11) Data acquisition and synchronization: software synchronization is adopted to synchronize various driving data by collecting CPS clock or vehicle terminal system clock to construct driving data including time, image data, radar original data, integrated navigation data and vehicle dynamics parameters structure; (12) Data encoding and caching: The image data is encoded by H264 or H265, and other data is encoded by virtual CAN information; data registers are set to cache the driving data structure. 3. L3 level automatic driving system road driving data collection, analysis and uploading method according to claim 1, is characterized in that described step 2. comprises the following steps: (21) automatic driving system intermediate result output interface definition; (22 ) target matching consistency detection; (23) localized road sign semantic output; (24) extreme vehicle operation detection; (25) human-machine decision consistency detection. 4. L3 level automatic driving system road driving data collection, analysis and uploading method according to claim 3, is characterized in that the concrete method of described step (22) target matching consistency detection is: According to the output results of the millimeter-wave radar and the visual perception system, extract the driving data whose timing matching distance tolerance is greater than the preset threshold dmin, and calculate the matching distance according to the following formula: in, output target coordinates for the radar, Output the target coordinates for the on-board camera, n is the target life cycle; if d>dmin, upload the driving data of the corresponding segment. 5. L3 level automatic driving system road driving data collection, analysis and uploading method according to claim 3, it is characterized in that the concrete method of described step (24) extreme vehicle operation detection is: The yaw rate, longitudinal acceleration, longitudinal deceleration and lateral acceleration output by the inertial navigation system are time-series encoded, and the time-series encoded data are classified into extreme vehicle operations by numerical analysis methods or machine learning methods. Extreme vehicle operations are classified into rapid acceleration operations. , sharp deceleration operation and sharp turning operation; The numerical analysis method is as follows: if the measured value of longitudinal acceleration is continuously greater than the set threshold value A2min for more than N times, it is confirmed as a rapid acceleration operation; if the measured value of longitudinal deceleration is continuously less than the set threshold value A2min for more than N times, it is confirmed as a rapid deceleration operation ; If the measured value of lateral acceleration and the measured value of yaw rate are continuously greater than the set threshold AYmin and Tmin for more than N times, it is confirmed as a sharp turning operation; The machine learning method is as follows: using extreme vehicle operation and driving time series data samples, offline training support vector machine or long short-term memory network; deploying the trained model above in the vehicle analysis terminal, the input is time-series encoded inertial navigation data, and the output is rapid acceleration operation. , an event signal for a sharp deceleration operation or a sharp turn operation. 6. L3 level automatic driving system road driving data collection, analysis and uploading method according to claim 3, it is characterized in that the concrete method of described step (25) man-machine decision consistency detection is: According to the output result of the planning layer and the output result of the vehicle pose, according to the preset preview distance, extract the driving data segment whose deviation between the actual trajectory and the planned trajectory is greater than the preset threshold Dmin; calculate the normalization in the vehicle coordinate system according to the following formula Trajectory deviation: Among them, [Xi, Yi] is the actual track point coordinates in the vehicle coordinate system, [xi, yi] is the planned track point coordinates in the vehicle coordinate system; m is the number of track points; if D>Dmin, upload the driving data of the corresponding segment. 7. The L3-level automatic driving system road driving data collection, analysis and uploading method according to claim 3 or 4 or 5 or 6, wherein the step (21) automatic driving system intermediate result output interface defines the specific The method is: including perception target output, location landmark semantic output, vehicle pose output and planning decision output; Among them, the perception target output: including the target output of the vision system, the target output of the millimeter wave radar system, and the original target point cloud output of the millimeter wave radar system; Semantic output of localized road signs: Output the semantics of localized road signs in the form of binary layers, including the semantic output of drivable areas, the semantic output of lane boundaries and the semantic output of indicative road signs; Vehicle pose output: including vehicle position, vehicle speed, heading angle and 6-axis inertial sensor output; Planning decision output: The preview point trajectory is given at fixed longitudinal space intervals. 8. The L3-level automatic driving system road driving data collection, analysis and upload method according to claim 3 or 4 or 5 or 6, wherein the concrete method of the step (23) locating the semantic output of road signs is: According to the location semantic output of the perception module, and use the vehicle pose estimation result of the location module to construct the location landmark semantic output; the location landmark semantic output method adopts the key frame semantic output method or the compressed full semantic output method; The key frame semantic output method is: according to the mileage integral estimation of the positioning module, extract a frame of key semantic output every 50 meters, and construct a key frame semantic location register. The key frame semantic location register stores the key frame and the vehicle latitude and longitude data at the corresponding time; Packets are compressed every 20 frames, and an upload request signal is sent; The compressed full-semantic output method is: including lane-level semantics and guidance-level semantics; post-processing to extract the number of lanes, lane widths, boundary types, lanes, and off-center distances from the visual system lane and drivable area semantic output layers; The post-processing of the road sign semantic output layer extracts two parts of the road sign and the space road sign; the above data and the latitude and longitude data of the vehicle at the corresponding time are compressed once every 1km traveled, and an upload request signal is sent. 9. The L3-level automatic driving system road driving data collection, analysis and upload method according to claim 1 or 2 or 3, characterized in that: The specific method of the data communication in the step (3) is: according to the online data analysis result in the step (2), compress the driving data queue obtained after the collection in the step (1), and name the corresponding compressed file according to a predefined rule; use the TCP or UDP protocol. , transparently transmit the compressed data to the server through the 4G network or wireless network; The specific method for receiving and storing the vehicle-side driving data in the step (4) is as follows: on the server side, through the TCP or UDP protocol, the driving data sent by the vehicle-mounted terminal is received; the data collection date is used as a subdirectory to store the relevant driving data.