CN114740866A - Robot autonomous exploration method and system based on deep learning - Google Patents

Abstract

The deep learning-based robot autonomous exploration method and system disclosed in the present disclosure include: acquiring robot coordinate points and a constructed current indoor scene graph; determining the boundary of the current indoor scene graph through a deep learning network; selecting the boundary center closest to the robot coordinate point is a candidate point; a tree-like path is generated between the robot coordinate point and the candidate point through the fast-growing random tree algorithm; the penultimate node on the tree-like path is selected as the target point; the global path is generated according to the robot coordinate point and the target point; Drive the robot to move along the global path at a set speed to obtain new laser point cloud data; update the current indoor scene graph through the new laser point cloud data to complete the indoor scene mapping. Realize high-precision indoor scene mapping.

Description

Robot autonomous exploration method and system based on deep learning

technical field

The present invention relates to the field of robotics, in particular to a method and system for autonomous exploration of robots based on deep learning.

Background technique

The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

Mobile robots have the advantages of flexible action, convenient operation and strong robustness, and have broad application prospects in medical, military, aerospace, logistics and other fields. Autonomous exploration of mobile robots means that robots in an unknown environment, through autonomous movement and environmental perception, finally build a complete environmental map. Robotic autonomous exploration can help people achieve map reconstruction of complex terrain, greatly reducing extreme environmental conditions. (such as small size, high temperature, etc.) adversely affect people, which provides convenience for people to operate the environmental map in the later stage.

In the existing technical solutions, the following technical difficulties exist when robots autonomously explore and map scenes:

(1) The exploration efficiency is low. When mapping a large area, extracting the map boundary requires a lot of computing resources, and it is difficult to obtain a boundary with a large amount of information, resulting in low efficiency of robot exploration.

(2) The map integrity is low. When the robot performs exploratory mapping, the final drawn environment map is incomplete due to the limitation of the target point selection area.

(3) The flexibility of mapping is poor. When the robot selects the target point, it is easy to select the point that is too close to the obstacle, which makes it difficult for the robot to reach the target point.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present disclosure proposes a method and system for autonomous exploration of robots based on deep learning. The U-net deep learning network is used to quickly determine the boundaries of known and unknown areas in the map, and candidate points are selected from the boundaries, and then through the U-net deep learning network. The fast growing random tree algorithm generates a tree-like path between the robot coordinate points and the candidate points, and then determines the target point according to the tree-like path. Interior scene mapping.

To achieve the above object, the present disclosure adopts the following technical solutions:

In the first aspect, a deep learning-based autonomous exploration method for robots is proposed, including:

Obtain the coordinate points of the robot and the current indoor scene graph constructed;

Determine the boundary of the current indoor scene graph through a deep learning network;

Select the boundary center closest to the robot coordinate point as the candidate point;

Generate tree-like paths between robot coordinate points and candidate points through fast growing random tree algorithm;

Select the penultimate node on the tree path as the target point;

Generate a global path based on robot coordinate points and target points;

Drive the robot to move along the global path at a set speed to acquire new laser point cloud data;

The current indoor scene map is updated through the new laser point cloud data to complete the indoor scene mapping.

In the second aspect, a robot autonomous exploration system based on deep learning is proposed, including:

The data acquisition module is used to acquire the coordinate points of the robot and the current indoor scene graph constructed;

The target selection module determines the boundary of the current indoor scene graph through the deep learning network; selects the boundary center closest to the robot coordinate point as the candidate point; generates a tree-like path between the robot coordinate point and the candidate point through the fast growing random tree algorithm; select The penultimate node on the tree path is the target point;

The path planning module is used to generate a global path according to the robot coordinate points and target points;

The robot drive module is used to drive the robot to move along the global path at a set speed to obtain new laser point cloud data;

The mapping module is used to update the current indoor scene map through the laser point cloud data to complete the indoor scene mapping.

In a third aspect, an electronic device is proposed, including a memory and a processor, and computer instructions stored in the memory and running on the processor, the computer instructions are executed by the processor to complete a deep learning-based robot autonomous exploration method the steps described.

In a fourth aspect, a computer-readable storage medium is provided for storing computer instructions, and when the computer instructions are executed by a processor, the steps described in the deep learning-based robot autonomous exploration method are completed.

Compared with the prior art, the beneficial effects of the present disclosure are:

1. According to the principle of edge detection, the present disclosure quickly extracts the boundary between the known area and the unknown area in the map through the U-net segmentation network, which helps to improve the speed of boundary extraction, obtains the optimal target point, and effectively solves the information of the target point. Small volume and regional limitations.

2. The present disclosure establishes a random tree between robot coordinate points and target points, which can simultaneously solve the problems of global path planning and target point selection.

3. The present disclosure adopts the RRT algorithm to select the target point, which effectively solves the problem that the target point is too close to the obstacle, so that the robot cannot reach it.

Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will become apparent from the description which follows, or may be learned by practice of the invention.

Description of drawings

The accompanying drawings that form a part of the present application are used to provide further understanding of the present application, and the schematic embodiments and descriptions of the present application are used to explain the present application and do not constitute improper limitations on the present application.

1 is a frame diagram of the method disclosed in Embodiment 1;

Fig. 2 is the map grid cell diagram with the center as the boundary cell disclosed in embodiment 1;

Fig. 3 is a map grid cell diagram with the center being a non-boundary cell disclosed in Embodiment 1;

4 is a structural diagram of the U-net deep learning network disclosed in Embodiment 1;

Fig. 5 is the RRT algorithm flow chart disclosed in embodiment 1;

6 is a schematic diagram of the robot selecting target point disclosed in Embodiment 1;

7 is a schematic diagram of the motion of the robot at adjacent moments disclosed in Embodiment 1;

8 is a schematic diagram of the robot speed sampling disclosed in Embodiment 1;

FIG. 9 is an architecture diagram of the Cartographer-SLAM algorithm disclosed in Embodiment 1. FIG.

Among them: 1. Occupied unit, 2. Free unit, 3. Unknown unit, 4. Boundary unit, 5. Non-boundary unit, 6. Fully connected layer, 7. Pooling layer, 8. Convolutional layer, 9. Map data Flow, 10, Obstacles, 11, Expansion Zones for Obstacles, 12, Robot Base, 13, Target Points, 14, Candidate Points, 15, Global Path, 16, Robot.

Detailed ways:

The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

In this disclosure, terms such as "top", "bottom", "left", "right", "front", "rear", "vertical", "horizontal", "side", "bottom", etc. The orientation or positional relationship is based on the orientation or positional relationship shown in the drawings, and is only a relational word determined for the convenience of describing the structural relationship of each component or element of the present disclosure, and does not specifically refer to any component or element in the present disclosure, and should not be construed as a reference to the present disclosure. public restrictions.

In the present disclosure, terms such as "fixed connection", "connected", "connected", etc. should be understood in a broad sense, indicating that it may be a fixed connection, an integral connection or a detachable connection; it may be directly connected, or through an intermediate connection. media are indirectly connected. For the relevant scientific research or technical personnel in the field, the specific meanings of the above terms in the present disclosure can be determined according to specific circumstances, and should not be construed as limitations on the present disclosure.

Example 1

In order to achieve high-precision indoor scene mapping, a deep learning-based robot autonomous exploration method is disclosed in this embodiment, including:

Obtain the coordinate points of the robot and the current indoor scene graph constructed;

Determine the boundary of the current indoor scene graph through a deep learning network;

Select the boundary center closest to the robot coordinate point as the candidate point;

Generate tree-like paths between robot coordinate points and candidate points through fast growing random tree algorithm;

Select the penultimate node on the tree path as the target point;

Generate a global path based on robot coordinate points and target points;

Drive the robot to move along the global path at a set speed to acquire new laser point cloud data;

The current indoor scene map is updated through the new laser point cloud data to complete the indoor scene mapping.

Further, the specific process of constructing the current indoor scene graph is as follows:

Obtain the laser point cloud data and the pose data of the robot at the current moment;

Build a subgraph according to the pose data of the robot;

The constructed subgraph is updated through the laser point cloud data at the current moment to obtain the current indoor scene graph.

Further, the pose data of the robot includes robot coordinate points, mileage data, attitude data and acceleration.

Further, the constructed current indoor scene graph is a two-dimensional occupied grid map, and the map grid cells are divided into: idle cells, unknown cells and occupied cells according to the grid values. When there is at least one unknown cell in the adjacent cells, the cell is a boundary cell, and all the boundary cells form a boundary.

Further, the specific process of generating a tree-like path between the robot coordinate point and the candidate point through the fast growing random tree algorithm is as follows:

Randomly select a point in the current indoor scene graph as a sampling point;

Select the node closest to the sampling point as the growth node;

Generate a new node from the growing node towards the sampling point;

Add new nodes that meet the conditions to the tree to generate a tree-like path.

Further, the selection process of the robot set speed is as follows:

Get the movement speed range of the robot;

Sampling multiple sets of speeds within the motion speed interval;

Calculate the trajectory of the robot at each speed;

Select the optimal motion trajectory from the robot motion trajectory through the evaluation function;

The speed corresponding to the optimal motion trajectory is the set speed for driving the robot to move.

Further, the motion of the robot is simplified to linear motion, and the motion model of the robot is constructed;

According to the motion model of the robot, the trajectory of the robot at each speed is calculated.

The deep learning-based robot autonomous exploration method disclosed in this embodiment is discussed in detail.

The robot autonomous exploration method based on deep learning, as shown in Figure 1, includes:

S1: Obtain the robot coordinate points and the constructed current indoor scene graph.

In the specific implementation, as shown in Figure 9, the Cartographer-SLAM mapping algorithm is used to construct the indoor scene graph. The SLAM algorithm is a SLAM algorithm based on graph optimization. The algorithm mainly consists of two parts: Local SLAM and Global SLAM. constitute. Among them, Local SLAM establishes a submap (Submap) by receiving the filtered adjacent frames of point cloud data. Whenever a frame of point cloud data is obtained, it is matched with the recently established submap, and the point cloud of the current frame is matched. The data is inserted into the subgraph (that is, the scan_match ing process), and the subgraph is updated by continuously inserting new point cloud data frames. The subgraph and pose estimation established by point cloud matching are reliable in a short time, but will generate a large cumulative error for a long time, so it needs to be further determined through loop closure detection and back-end optimization (G l obal SLAM). The global pose and global map of the robot.

Among them, the specific process of constructing the current indoor scene graph is as follows:

Obtain the laser point cloud data and the pose data of the robot at the current moment, where the pose data of the robot includes sensory data such as robot coordinate points, mileage data, attitude data and acceleration;

Build a subgraph according to the pose data of the robot;

The constructed subgraph is updated through the laser point cloud data at the current moment to obtain the current indoor scene graph.

The indoor scene map constructed in this embodiment adopts a two-dimensional occupancy grid map.

S2: Determine the boundary of the current indoor scene graph, and select the boundary center closest to the robot coordinate point as the candidate point; generate a tree-like path between the robot coordinate point and the candidate point through the fast-growing random tree algorithm; select the reciprocal on the tree-like path The second node is the target point.

In specific implementation, the target selection module includes a frontier detector (Frontier Detector) and a global random tree planner (Global Planner), which are responsible for extracting boundary points in the map and selecting an optimal one from the boundary points. The target point, which is used to build the global path of the robot later.

The boundary detector is used to quickly determine the boundary of the current indoor scene graph. Since the indoor scene graph constructed in this embodiment is a two-dimensional occupied grid map, the occupied grid map (Occupied Grid Map) is a grid map. There is a grid value, which is the probability value of the grid being occupied. Therefore, by filling the grid with different map values with different colors, a map representing the environmental characteristics can be obtained, as shown in Figure 2, according to the grid value. The map grid cells can be divided into free cells 2 (Free), unknown cells 3 (Unknown), and occupied cells 1 (Occupied). In this embodiment, the boundary cells are defined as grid sets that meet the following conditions: Condition 1: As shown in FIG. 3 , the boundary unit 4 is an idle unit; Condition 2: As shown in FIG. 2 , there is at least one unknown unit 3 in the eight adjacent grids of the boundary unit 4 . A boundary is a collection of boundary elements. Fast extraction of boundaries in maps by using U-net deep learning network. The reason for extracting the boundaries of the map is that the boundary areas of the map contain a large amount of unknown information, and you can obtain greater information gain by going to these unknown areas.

The structure of the U-net deep learning network is shown in Figure 4, including pooling layer 7, convolution layer 8 and fully connected layer 6, map data stream 9 in pooling layer 7, convolution layer 8 and fully connected layer 6 Circulation, to determine the boundaries of the map.

The global planner (Global Planner) is used to select the target point that the robot is currently going to and generate the global path from the robot coordinate point to the target point.

Among them, the specific process of determining the target point is as follows:

Select the boundary center closest to the robot coordinate point as the candidate point that the robot is currently going to;

Generate tree-like paths between robot coordinate points and candidate points through fast growing random tree algorithm;

Select the penultimate node on the tree path as the target point.

The rapid growing random tree algorithm (RRT algorithm) is a tree-like algorithm based on sampling, as shown in Figure 4, by randomly sampling the points on the map as sampling points, selecting the node closest to the sampling point as the growing node, and then growing The node generates a new node in the direction of the sampling point, and adds a new node that meets certain conditions to a part of the tree.

As shown in Figure 5, the rules for random sampling points of RRT are: randomly select a point in the specified space and take the target point as the sampling point alternately. For example, if the sampling point of this RRT is randomly obtained in a limited space , the sampling point of the next RRT will be selected at the candidate point, and repeated iterations in this way can ensure that the RRT will eventually converge to the candidate point 14 .

When the penultimate node on the tree path is selected as the target point 13, and then the global path 15 is determined according to the target point, the robot 16 can be prevented from entering the expansion area 11 of the obstacle, and the robot base 12 can be prevented from being too close to the obstacle 10.

S3: Generate a global path based on robot coordinate points and target points.

During specific implementation, a global path is generated by a path planning module, and the path planning module includes a global planner (Global Planner) and a local planner (Local Planner).

Among them, after selecting the final target point, the global planner determines the global path according to the coordinate points and target points of the robot. In the local planning period, a local planning algorithm based on dynamic windows (DWA algorithm) is used to make the robot move along the global path while moving along the global path. , which can avoid nearby obstacles in real time. The main idea of the DWA algorithm is to sample multiple sets of speeds in the robot's speed space (v, w), and then simulate the robot in a certain time interval (sim_per i od) at these speeds. After obtaining multiple sets of motion trajectories, these trajectories are evaluated through the evaluation function, and the speed (v_best, w_best) corresponding to an optimal trajectory is selected to drive the robot to move.

The process of calculating the motion trajectory of the robot is as follows: Assuming that the robot moves in a plane and the robot does not move omnidirectionally, the motion of the robot can be simplified as a linear motion in adjacent moments. Among them, the motion model of the robot's coordinates and yaw angle after a period of time is:

Among them, x _t , y _t , and θ _t are the position coordinates and yaw angle of the robot after moving for a period of time, respectively, x, y, and θ are the current position coordinates and yaw angle of the robot, respectively, and v and w are the current position coordinates and yaw angle of the robot, respectively. Linear velocity and angular velocity, Δt is the movement time interval, α is the angle between the position vector and the positive half-axis of the x-axis before and after the robot moves, as shown in Figure 7.

As shown in Figure 8, during the actual motion of the robot, its speed will be subject to various constraints, such as the maximum speed of the motor, the torque of the motor, the maximum braking distance, and the safety distance from obstacles, etc. Based on the above constraints, it can be obtained in this At the moment, the interval of the robot's motion speed, and then within this interval, multiple groups of speeds are uniformly sampled, and the robot's speed remains unchanged within a certain time interval (sim_period).

After obtaining a set of velocities and their corresponding predicted motion trajectories through velocity sampling, the following evaluation functions are used to evaluate multiple motion trajectories:

G(v, w)=w ₁ Δθ(Δt)+w ₂ D _nearest +w ₃ v (2)

Among them, w _i is the weight coefficient, Δθ(Δt) is the yaw angle change of the robot in a period of time, D _nearest is the distance between the robot and the nearest obstacle, and v is the current speed of the robot. Through the evaluation function, a path that maximizes G(v, w) is selected as the optimal motion path, and the corresponding speed is selected to drive the robot to complete the local path planning.

S4: Drive the robot to move along the global path through the speed corresponding to the selected optimal motion trajectory to obtain new laser point cloud data; update the current indoor scene graph through the new laser point cloud data to complete indoor scene mapping.

The method disclosed in this embodiment can quickly extract the boundary between the known area and the unknown area in the map through the deep learning method, select candidate points, and then generate a tree shape between the robot coordinate points and the candidate points through the fast growing random tree algorithm path, and further select the penultimate node on the tree path as the target point, and finally generate a global path between the robot and the target point, which effectively solves the problem that the target point is too close to the obstacle and the robot cannot reach it. By establishing a random tree between the robot coordinates and the target point, the problem of global path planning and target point selection can be solved at the same time. According to the principle of edge detection, the boundary between the known area and the unknown area in the map can be extracted, which helps to obtain the optimal target. It effectively solves the problems of small amount of information and regional limitations of target points. The Cartographer-SLAM mapping algorithm used can perform loop closure detection through matching between laser point clouds and sub-images, which can achieve high-precision indoor scene mapping.

Example 2

In this embodiment, a robot autonomous exploration system based on deep learning is disclosed, including:

The data acquisition module is used to acquire the coordinate points of the robot and the current indoor scene graph constructed;

The target selection module determines the boundary of the current indoor scene graph through the deep learning network; selects the boundary center closest to the robot coordinate point as the candidate point; generates a tree-like path between the robot coordinate point and the candidate point through the fast growing random tree algorithm; select The penultimate node on the tree path is the target point;

The path planning module is used to generate a global path according to the robot coordinate points and target points;

The robot drive module is used to drive the robot to move along the global path at a set speed to obtain new laser point cloud data;

The mapping module is used to update the current indoor scene graph through the new laser point cloud data to complete the indoor scene mapping.

Example 3

In this embodiment, an electronic device is disclosed, which includes a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the based on Steps described in a deep learning approach to autonomous robotic exploration.

Example 4

In this embodiment, a computer-readable storage medium is disclosed, which is used to store computer instructions. When the computer instructions are executed by a processor, the steps described in the deep learning-based robot autonomous exploration method disclosed in Embodiment 1 are completed. .

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the present invention can still be Modifications or equivalent replacements are made to the specific embodiments of the present invention, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (10)

1. The robot autonomous exploration method based on deep learning is characterized by comprising the following steps:

acquiring a coordinate point of the robot and a constructed current indoor scene graph;

determining the boundary of the current indoor scene graph through a deep learning network;

selecting a boundary center closest to a coordinate point of the robot as a candidate point;

generating a tree-shaped path between the robot coordinate point and the candidate point by a fast growing random tree algorithm;

selecting the penultimate node on the tree-shaped path as a target point;

generating a global path according to the robot coordinate points and the target point;

driving the robot to move along the global path at a set speed to acquire new laser point cloud data;

and updating the current indoor scene graph through the new laser point cloud data to complete indoor scene graph building.

2. The deep learning-based robot autonomous exploration method according to claim 1, wherein a specific process of constructing a current indoor scene graph is as follows:

acquiring laser point cloud data of the current moment and pose data of the robot;

constructing subgraphs according to the pose data of the robot;

and updating the constructed subgraph through the laser point cloud data at the current moment to obtain a current indoor scene graph.

3. The deep learning-based robot autonomous exploration method according to claim 2, characterized in that pose data of the robot comprises robot coordinate points, mileage data, pose data and acceleration.

4. The deep learning-based robot autonomous exploration method according to claim 1, characterized in that the constructed current indoor scene graph is a two-dimensional occupancy grid map, and the grid cells of the map are divided into: the cell comprises a free cell, an unknown cell and an occupied cell, wherein when one cell is the known cell and at least one unknown cell is in eight adjacent cells of the cell, the cell is a boundary cell, and all boundary cells form a boundary.

5. The deep learning-based robot autonomous exploration method according to claim 1, characterized in that the specific process of generating a tree-like path between a robot coordinate point and a candidate point by a fast growing random tree algorithm is as follows:

randomly selecting a point in a current indoor scene graph as a sampling point;

selecting a node closest to the sampling point as a growth node;

generating a new node from the growing node towards the direction of the sampling point;

and adding the new nodes meeting the conditions into the tree to generate a tree-shaped path.

6. The deep learning-based robot autonomous exploration method according to claim 1, wherein the robot set speed is selected by the following process:

acquiring a motion speed interval of the robot;

sampling a plurality of groups of speeds in a movement speed interval;

calculating the motion track of the robot at each group of speeds;

selecting an optimal motion track from the motion tracks of the robot through an evaluation function;

the speed corresponding to the optimal motion track is the set speed for driving the robot to move.

7. The deep learning-based robot autonomous exploration method according to claim 6, wherein robot motion is simplified into linear motion, and a motion model of the robot is constructed;

and calculating the motion trail of the robot at each group of speeds according to the motion model of the robot.

8. Robot is system of independently exploring based on degree of depth learning, its characterized in that includes:

the data acquisition module is used for acquiring coordinate points of the robot and a constructed current indoor scene graph;

the target selection module is used for determining the boundary of the current indoor scene graph through a deep learning network; selecting a boundary center closest to a coordinate point of the robot as a candidate point; generating a tree-shaped path between the robot coordinate point and the candidate point by a fast growing random tree algorithm; selecting the penultimate node on the tree-shaped path as a target point;

the path planning module is used for generating a global path according to the robot coordinate points and the target point;

the robot driving module is used for driving the robot to move along the global path at a set speed to acquire new laser point cloud data;

and the mapping module is used for updating the current indoor scene map through the new laser point cloud data to complete indoor scene mapping.

9. An electronic device comprising a memory and a processor, and computer instructions stored on the memory and executed on the processor, the computer instructions, when executed by the processor, performing the steps of the deep learning based robotic autonomous exploration method of any of claims 1-7.

10. A computer-readable storage medium storing computer instructions which, when executed by a processor, perform the steps of the deep learning based robotic autonomous exploration method of any of claims 1-7.

Images