Vehicle beyond visual range real-time sensing method based on laser radar vehicle-road cooperation

Abstract

The invention relates to the field of intelligent network connection automobile sensing, in particular to a vehicle beyond visual range real-time sensing method based on laser radar vehicle-road cooperation. The method comprises the following steps that a roadside radar and a vehicle send point cloud information to a roadside unit RSU in a unified mode; the roadside unit aligns the vehicle point cloud with the roadside point cloud, determines the central position of the vehicle point cloud, and fuses the plurality of point clouds into an integral fused point cloud; the RSU divides the overall target detection task of the fusion point cloud into a plurality of small tasks, the small tasks are distributed to the computing nodes, and after the computing nodes complete the distributed tasks, results are returned to the RSU; the road side unit summarizes the calculation results of all the nodes to obtain a target detection result of the fused point cloud, and broadcasts the detection result to the vehicle, and the method improves the overall perception range of the vehicle, and can realize the overall beyond-visual-range perception of the road for each vehicle.

Description

Technical Field

[0001] The present invention relates to the field of intelligent connected vehicle perception, and specifically to a vehicle beyond-visual-range real-time perception method based on laser radar vehicle-road collaboration. Technical Background

[0002] With the increasing popularity of automobiles, automobiles have become a common means of transportation. In order to meet user needs, the development direction of automobiles has gradually shifted towards intelligence. V2X (Vehicle to Everything) technology is one of the supporting technologies in the field of smart cars and smart transportation. It includes the interaction between V2V (Vehicle to Vehicle) between vehicles and roads, V2I (Vehicle to Infrastructure) between vehicles and people, and V2N (Vehicle to Network) between vehicles and networks. The V2X technology system provides a Uu interface based on supporting large bandwidth and wide coverage communication services between terminals and base stations, and a PC5 interface based on low latency and high reliability requirements between terminals. Specifically, people, vehicles, and road infrastructure can communicate directly without passing through network base stations through the PC5 interface, while vehicles can communicate with network base stations or cloud nodes through the Uu interface. In general V2X road scenarios, roadside units RSU (Road Side Unit) will be installed on the roadside, and on-board units OBU (On Board Unit) will be installed on vehicles. In the specific implementation, the communication between vehicles (V2V) is realized by the message transmission between the on-board units (OBU) of different vehicles through the PC5 port; the communication between vehicles and road facilities (V2I) is realized by the message transmission between the on-board unit (OBU) of the vehicle and the roadside unit (RSU) of the roadside through the PC5 port. If it is within the coverage of the cellular mobile network, the OBU and RSU can also transmit messages with the network base station through the Uu port to obtain richer and more accurate information;

[0003] V2X technology combines the Uu interface and the PC5 interface to support each other and work together on information transmission, thereby ensuring the reliability of vehicle network communication with effective redundancy. Regardless of whether it is within the coverage of the mobile communication network, V2X technology can stably realize vehicle-road cooperative communication and ensure the stability and reliability of the system.

[0004] On the other hand, in the field of intelligent connected vehicles, environmental perception is one of the key parts of the autonomous driving system. The field of view coverage of environmental perception, the accuracy of target detection, and the total delay of task completion are directly related to the safety of autonomous driving. However, currently, intelligent connected vehicles mainly rely on sensors such as cameras, lidar, and millimeter-wave radars installed on a single vehicle for perception. This method of relying solely on single-vehicle perception cannot be applied to scenes with single-vehicle perception blind spots such as intersections and obstacles, and is very likely to cause traffic accidents. At the same time, due to the limitations of on-board computing resources, when the road traffic is heavy, the data from multi-source sensors is difficult to be processed quickly and effectively, the system's response speed is slow, and it is impossible to provide timely and accurate feedback to the driver in an emergency. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a vehicle beyond-visual-range real-time perception method based on lidar vehicle-road collaboration, so as to improve the overall perception range of the vehicle and realize beyond-visual-range perception of the entire road for each vehicle.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A vehicle beyond-visual-range real-time perception method based on laser radar vehicle-road collaboration includes the following steps:

[0008] S1, the vehicle uses the on-board radar to obtain the point cloud information around the vehicle, the roadside radar obtains the point cloud information on the roadside, and the roadside radar and the vehicle send the point cloud information to the roadside unit RSU;

[0009] S2, the roadside unit RSU transforms the coordinates of the vehicle point cloud based on the vehicle offset information recorded by the vehicle inertial sensor IMU to align it with the roadside point cloud; determines the center position of the vehicle point cloud based on the vehicle position information, and fuses multiple point clouds into an overall fused point cloud;

[0010] S3, the roadside unit RSU gives the best task allocation scheme based on the adaptive task offloading algorithm with the goal of minimizing the total system delay, divides the overall fusion point cloud target detection task into multiple small tasks, and allocates them to each computing node. Among them, the computing node is mainly the roadside unit RSU, and the vehicles on the road are supplemented by multiple nodes with computing capabilities. After each computing node completes the assigned task, it returns the result to the roadside unit;

[0011] S4, the roadside unit summarizes the calculation results of each node, obtains the target detection result of the fused point cloud, and broadcasts the detection result to the vehicle.

[0012] Furthermore, it also includes S5, when the detection result shows that there is an obstacle within the warning range set by the vehicle, a warning is promptly given to the driver through the vehicle system.

[0013] Furthermore, in step S1, the vehicle uses the octree algorithm to losslessly compress the original point cloud locally, and sends the compressed point cloud to the roadside unit RSU; in step S2, the RSU restores the compressed point cloud to the original point cloud through the corresponding decompression algorithm.

[0014] Furthermore, in step 2, the fused point cloud is obtained in the following manner:

[0015] Point Cloud i For vehicle v i Point cloud, i = 1, 2, ... N, N represents the number of vehicles, the roadside point cloud is recorded as P0:

[0016] (1) Vehicle v i During the communication with the roadside unit RSU, it always shares its own positioning information and the vehicle offset information R recorded by the IMU i , where the yaw angle, pitch angle and roll angle are α i , β i and γ i ,have:

[0017] R i =R z (α i )R y (β i )R x (γ i )

[0018]

[0019]

[0020] (2) Vehicle point cloud P i Transform the coordinates so that they align with the coordinates of P0:

[0021] According to vehicle v i The position information of the roadside radar and the known position information of the roadside radar are used to calculate the point cloud P i Compared with the center point position offset Δd of point cloud P0 xi , Δd yi and Δd zi ; Point cloud P i Expressed as The transformed point cloud P′ i Expressed as

[0022]

[0023] (3) All transformed vehicle point clouds P′ i , i = 1, 2, ... N, and merge with point cloud P0 to obtain a new fused point cloud P f ,have:

[0025]

[0026] Furthermore, the adaptive task offloading algorithm in step 3 includes:

[0027] Roadside unit RSU fusion point cloud P f Perform uniform redundant division to obtain multiple small point clouds P f1 , P f2 ..., the detection of a small point cloud is a subtask q k ; There are n available computing nodes on the current road, then the fused point cloud P f The redundancy is divided into n small point clouds, and the corresponding detection task Q is divided into n subtasks, Q = {q1, q2, ...q k …,q n};

[0028] When assigning nodes to the subtasks of the detection task Q, a node only performs one subtask at a time; for subtask q k ∈Q, the set of available auxiliary computing nodes is recorded as At the same time, the subtask q k Offloading to Node Recorded as action a k,i , subtask q k The action space A k The statement is as follows:

[0029]

[0030] Taking the point cloud of consecutive T frames as a cycle, the detection of the first M, n≤M≤T, frame point clouds in this cycle is the initial exploration stage of the algorithm. This stage will explore each action a in the action space k,i , and record each action a k,i The completion delay t k,i ; After the initial exploration phase is completed, the different subtasks are unloaded q k To different nodes b i The task completion delay;

[0031] After the initial exploration phase is completed, subtask q is performed frame by frame iteratively. k Uninstall action selection, where action a k,i The completion delay t k,iIt is estimated as the mean completion time of the same action in all frames before the current frame, and the subtask q is recorded as k The lower bound of the delay of offloading to different nodes is according to Unload the subtasks in order from large to small, that is The largest subtasks are unloaded first. The smallest subtask is uninstalled last;

[0032] For each subtask q k , select the action a with the highest score W k,i Execution, i.e. subtask q k Uninstall to a k,i The corresponding node b i , specifically:

[0033] For subtask q k ∈Q, each action a k,i The reward is expressed as in Represents subtask q k The upper bound of the delay when unloading to different nodes, y = U (x) is a simple linear function;

[0034] At the sth choice, action a k,i The number of times selected is recorded as S k,i (s), action a k,i The uncertainty measure δ of is expressed as follows:

[0035]

[0036] The score W consists of the expected reward u and an uncertainty measure, namely:

[0037]

[0038] Among them, ξ represents the exploration rate;

[0039] Unload each subtask of the detection task Q to the current best node in turn. After waiting for the subtask to be completed and the parameters to be updated, the detection task of the next frame of point cloud can be carried out.

[0040] After completing the detection task of T frame point clouds in one cycle, the algorithm is initialized and the next cycle begins.

[0041] Furthermore, ξ is set to λ is the current iteration number.

[0042] Furthermore, each computing node uses the pre-trained PointPillars model to complete the target detection task.

[0043] The present invention is based on advanced 3D point cloud target detection technology and vehicle network communication technology, fully utilizes vehicle-mounted lidar sensors and roadside computing units, and fuses point clouds on the road through early point cloud fusion technology, thereby improving the range and accuracy of target detection, overcoming the blind spot problem that may exist in single-vehicle perception, and enabling vehicles to perceive possible obstacles beyond the visual range in real time; fully utilizes the computing power of each node on the road, and unloads the detection task to multiple nodes through the task offloading algorithm, thereby speeding up task processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of a T-junction scene in an embodiment of the present invention.

[0045] Figure 2 This is a flowchart of the specific implementation steps of the vehicle beyond-visual-range real-time perception method based on lidar vehicle-road collaboration in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is further described below in conjunction with specific embodiments and drawings.

[0047] In this example, if Figure 1 As shown, the roadside unit RSU provides services for a T-junction, on which there are vehicles v1, v2 and roadside radar r0; vehicles v1 and v2 communicate with the roadside unit RSU through V2X, and the roadside radar communicates with the roadside unit RSU through a network cable. When vehicle v1 is driving on the road, due to the obstruction of obstacles (trucks in the figure), the camera and laser radar of vehicle v1 cannot perceive pedestrian 1 who is about to cross the road. At this time, through the vehicle beyond-visual-range real-time perception method based on laser radar vehicle-road collaboration designed by the present invention, vehicle v1 can share the perception data of vehicle v2 and roadside radar r0 on the road in real time, and know the approximate positions of pedestrians 1, pedestrian 2 and pedestrian 3 on the road, so as to achieve the beyond-visual-range perception effect. Furthermore, vehicle v1 can also calculate the approximate distance between each pedestrian and itself. If the distance is less than the preset safety alarm distance, the vehicle system can automatically issue a warning to the driver to reduce the possibility of traffic accidents.

[0048] In this example, vehicles v1 and v2 are equipped with an onboard unit OBU, and a roadside unit RSU is installed on the roadside. After the OBU and RSU are system burned and simply developed, they can be used to communicate through the PC5 port in the same frequency band regardless of whether they are in a cellular mobile network. Vehicles v1 and v2 communicate via V2V (OBU communicates with OBU), vehicle v1 communicates with the roadside unit via V2I (OBU communicates with RSU), and vehicle v2 communicates with the roadside unit via V2I (OBU communicates with RSU).

[0049] The vehicle beyond-visual-range real-time perception method based on LiDAR vehicle-road collaboration in this example includes the following steps:

[0050] After S1 vehicles v1 and v2 enter the signal coverage of the roadside unit RSU and successfully establish communication, vehicles v1 and v2 choose to share their perception data with the roadside unit. The roadside unit adds vehicles v1 and v2 to the service vehicle queue, as follows:

[0051] S1.1 Vehicle v1 obtains original point cloud information P1 around the vehicle through the on-board laser radar, vehicle v2 obtains original point cloud information P2 around the vehicle through the on-board laser radar, and the roadside radar obtains original point cloud information P0 around the roadside;

[0052] S1.2 performs lossless compression on the original point clouds P1, P2 and P0 locally, and the compression algorithm adopts the octree point cloud compression algorithm provided in the PCL point cloud library to obtain bit streams b1, b2 and b0;

[0053] S1.3 Vehicle v1 and vehicle v2 send bit streams b1 and b2 to the roadside unit RSU through V2X communication, and the roadside radar sends bit stream b0 to the roadside unit RSU through network cable communication.

[0054] S2 Roadside Unit RSU obtains road global environment perception data based on the perception data of each vehicle and roadside radar;

[0055] S2.1 After receiving the bit streams b1, b2 and b0, the roadside unit RSU decompresses and restores them to point clouds P1, P2 and P0;

[0056] S2.2 The roadside unit RSU transforms the coordinates of the vehicle point clouds P1 and P2 to align them with the roadside point cloud P0, and merges the three point clouds into the overall road point cloud P f ;

[0057] Take the coordinate transformation of point cloud P1 as an example:

[0058] (1) Vehicle v1 always shares its own positioning readings and the vehicle offset information R1 recorded by the IMU during communication with the roadside unit RSU, where the yaw angle, pitch angle, and roll angle are α1, β1, and γ1 respectively;

[0059] R1=R z (α1)R y (β1)R x (γ1)

[0060]

[0061] (2) Since the vehicle point cloud and the roadside point cloud are collected at different positions and angles, it is necessary to transform the coordinates of P1 so that it is aligned with the coordinates of P0. Based on the GPS positioning reading of vehicle v1 and the known roadside radar position information, the center point position offset Δd of point cloud P1 compared to point cloud P0 is calculated. x1 , Δd y1 and Δd z1 ; Point cloud P1 is represented as The transformed point cloud P′1 is expressed as

[0062]

[0063] (2) By fusing point cloud P′1 with point cloud P0, a new point cloud can be obtained, which expands the overall perception range of the point cloud:

[0064]

[0065] The coordinate transformation process of point cloud P2 is similar to that of P1. By analogy, after fusing all point cloud data, the final fused point cloud P is obtained. f .

[0066] S3 distributed computing detection tasks;

[0067] Since the fused point cloud is formed by the fusion of multiple point clouds and may contain a large amount of information, detection by a single computing node alone may cause a large delay. In order to make full use of the computing resources on the road and complete the target detection task of the fused point cloud in the shortest possible time, this method considers dividing the detection task of the fused point cloud into multiple subtasks and assigning them to multiple computing nodes for parallel computing to reduce the total delay of the system. The logic of the division of the detection task is: according to the number of current computing nodes on the road, the fused point cloud is redundantly divided, and the point cloud space is evenly divided into multiple small point clouds, which are handed over to different computing nodes for simultaneous detection. The target detection task of each small point cloud is a subtask. The benefit of having a certain redundancy between each small point cloud is that it can avoid information loss to a certain extent and facilitate the summary after the detection is completed. After each node independently detects its own task and summarizes it, the detection task of the entire point cloud can be completed. The computing node refers to a series of nodes with computing capabilities, mainly roadside units RSU and supplemented by vehicles on the road.

[0068] Specifically, the roadside unit RSU will fuse the point cloud P f Divided into multiple small point clouds P f1 , P f2 ...Detection of a small point cloud is a subtask q k, select the best task allocation scheme according to the adaptive task offloading algorithm, and offload the subtasks to each computing node. For example, in this embodiment, the optional computing nodes are vehicle v1, vehicle v2 and roadside unit RSU.

[0069] In this example, the decision of the task offloading algorithm is to offload small tasks, and the reward is to reduce the task delay. Each time the detection task is updated, the task offloading solution that can maximize the reward is selected. The specific task offloading algorithm used is an adaptive task offloading algorithm based on the multi-armed bandit MAB (Multi-Armed Bandit). Unlike the traditional greedy algorithm, this method considers the uncertainty factor measurement each time the node is selected, and will not blindly select the current optimal node, which can avoid the algorithm from falling into the local optimal situation to a certain extent. The MAB algorithm has three main components: arms, selection, and rewards. Specifically, a multi-armed bandit consists of multiple arms, and the decision of each arm will generate a reward. The MAB algorithm estimates the reward of each arm by balancing exploration and exploitation, and selects the best arm according to the current environment.

[0070] Without loss of generality, let n available computing nodes exist on the current road, and the detection task Q is divided into n subtasks, Q = {q1, q2, …, q n};

[0071] When assigning nodes to the subtasks of the detection task Q, a node can only perform one subtask at a time. k ∈Q, the set of available auxiliary computing nodes is recorded as At the same time, the subtask q k Offloading to Node Recorded as action a k,i , so the subtask q k The action space A k The statement is as follows:

[0072]

[0073] Taking the point cloud of consecutive T frames as a cycle, it can be assumed that the road scene will not change much in these consecutive T frames, that is, in this cycle, the point cloud detection subtask q of the same part after the point cloud of each frame is divided k The detection of the first N frames of point cloud in this cycle is the initial exploration stage of the algorithm, which explores each action a in the action space. k,i , and record each action a k,i The completion delay t k,i ,To ensure that each action is executed, the value of N needs to be greater than or equal to the number of subtasks n, that is, n≤N≤T.

[0074] In the initial exploration phase of the algorithm, the task completion delay of offloading different subtasks to different nodes can be obtained. After the exploration phase is completed, the subsequent point cloud offloading scheme is optimized. At this time, action a k,i The completion delay t k,i It can be estimated as the average completion time of the same action in all previous frames. In order to balance the detection delay of all subtasks and make each subtask complete in as close a time as possible, the subtask q k The lower bound of the delay of offloading to different nodes is according to Unload the subtasks in order from large to small, that is The largest subtasks are unloaded first. The smallest subtasks are uninstalled last.

[0075] For subtask q k ∈Q, each action a k,i The reward can be expressed as in Represents subtask q k The upper bound of the delay of unloading to different nodes, y = U (x) is a simple linear function. In the sth selection, action a k,i The number of times selected is recorded as S k,i (s). At the sth selection, action a k,i The number of times selected S k,i The more (s), the lower the uncertainty measure of the action. Based on the upper confidence bound algorithm of the multi-armed bandit, action a k,i The uncertainty measure δ is expressed as follows:

[0076]

[0077] The goal of each iteration of the algorithm is to select the unloading node with the maximum score for the subtask. The score W consists of the expected reward u and the uncertainty measure, namely:

[0078]

[0079] Among them, ξ represents the exploration rate, which is used to increase the uncertainty of the selection. Specifically, ξ is usually set to λ is the current number of iterations;

[0080] Unload each subtask of the detection task Q to the current best node in turn, wait for the subtask to be completed and update the delay t k,i After setting the parameters, the next frame point cloud detection task can be carried out.

[0081] After completing the detection task of T frame point clouds in one cycle, the algorithm is initialized and the next cycle begins.

[0082] Each computing node is responsible for using the pre-trained PointPillars model to complete the small point cloud target detection task. Prior to this, the 3D point cloud target detection model PointPillars is pre-trained using the KITTI dataset to obtain a model suitable for detecting road obstacles. The KITTI dataset annotates a variety of targets such as cars, trucks, and pedestrians, which can be used to train models for detecting road obstacles.

[0083] After each computing node completes a small task, it returns the result to the roadside unit RSU.

[0084] The S4 roadside unit summarizes the results and obtains the target detection results of the entire fused point cloud. In this embodiment, the detection results of pedestrian 1, pedestrian 2, and pedestrian 3 are obtained. The detection result of pedestrian 1 is: center point coordinates (x1, y1, z1), length, width, and height of the detection box (l1, w1, h1), category (Pedestrian), and confidence (score1); the detection result of pedestrian 2 is: center point coordinates (x2, y2, z2), length, width, and height of the detection box (l2, w2, h2), category (Pedestrian), and confidence (score2); the detection result of pedestrian 3 is: center point coordinates (x3, y3, z3), length, width, and height of the detection box (l3, w3, h3), category (Pedestrian), and confidence (score3). When the confidence score of the target is greater than the set threshold, the detection result of the target is considered valid;

[0085] Based on the position of the roadside radar and the vehicle's GPS reading, a simple numerical transformation is performed on the coordinates of each pedestrian to obtain the position of each pedestrian relative to the vehicle;

[0086] After completing the processing of the target detection results of the fused point cloud, the roadside unit RSU broadcasts the results to vehicles v1 and v2 in the service queue, and learns the locations of possible obstacles on the road. The vehicles can then obtain beyond-visual-range perception effects. When the detection results show that there are obstacles within the warning range set by the vehicle, the on-board system promptly issues a warning to the driver, thereby reducing the probability of traffic accidents and protecting the lives and property of drivers and pedestrians.

[0087] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions and / or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by the present invention shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A vehicle beyond visual range real-time perception method based on laser radar vehicle-road collaboration, characterized in that: The following steps are involved: S1, the vehicle uses the on-board radar to obtain the point cloud information around the vehicle, the roadside radar obtains the point cloud information on the roadside, and the roadside radar and the vehicle send the point cloud information to the roadside unit RSU; S2, the roadside unit RSU transforms the coordinates of the vehicle point cloud based on the vehicle offset information recorded by the vehicle inertial sensor IMU to align it with the roadside point cloud; determines the center position of the vehicle point cloud based on the vehicle position information, and fuses multiple point clouds into an overall fused point cloud; S3, the roadside unit RSU gives the best task allocation scheme based on the adaptive task offloading algorithm with the goal of minimizing the total system delay, divides the overall fusion point cloud target detection task into multiple small tasks, and allocates them to each computing node. Among them, the computing node is mainly the roadside unit RSU, and the vehicles on the road are supplemented by multiple nodes with computing capabilities. After each computing node completes the assigned task, it returns the result to the roadside unit; S4, the roadside unit summarizes the calculation results of each node, obtains the target detection result of the fused point cloud, and broadcasts the detection result to the vehicle.

2. The method according to claim 1, characterized in that It also includes S5, when the detection result shows that there is an obstacle within the warning range set by the vehicle, a warning is promptly given to the driver through the vehicle system.

3. The method according to claim 1, characterized in that In step S1, the vehicle uses the octree algorithm to losslessly compress the original point cloud locally, and sends the compressed point cloud to the roadside unit RSU; in step S2, the RSU restores the compressed point cloud to the original point cloud through the corresponding decompression algorithm.

4. The method according to claim 1, characterized in that: In step 2, the fused point cloud is obtained as follows: Point Cloud i For vehicle v i Point cloud, i = 1, 2, ... N, N represents the number of vehicles, the roadside point cloud is recorded as P0: (1) Vehicle v i During the communication with the roadside unit RSU, it always shares its own positioning information and the vehicle offset information R recorded by the IMU i , where the yaw angle, pitch angle and roll angle are α i , β i and γ i ,have: R i =R z (a i )R y (b i )R x (c i ) (2) Vehicle point cloud P i Transform the coordinates so that they align with the coordinates of P0: According to vehicle v i The position information of the roadside radar and the known position information of the roadside radar are used to calculate the point cloud P i The offset of the center point position compared to point cloud P0 and Point Cloud i Expressed as The transformed point cloud P′ i Expressed as (3) All transformed vehicle point clouds P′ i , i = 1, 2, ... N, and merge with point cloud P0 to obtain a new fused point cloud P f ,have:

5. The method according to claim 1, characterized in that The adaptive task offloading algorithm in step 3 includes: Roadside unit RSU fusion point cloud P f Perform uniform redundant division to obtain multiple small point clouds The detection of a small point cloud is a subtask q k ; There are n available computing nodes on the current road, then the fused point cloud P f The redundancy is divided into n small point clouds, and the corresponding detection task Q is divided into n subtasks, Q = {q1, q2, ...q k …,q n }; When assigning nodes to the subtasks of the detection task Q, a node only performs one subtask at a time; for subtask q k ∈Q, the set of available auxiliary computing nodes is recorded as At the same time, the subtask q k Offloading to Node Recorded as action a k,i , subtask q k The action space A k The statement is as follows: Taking the point cloud of consecutive T frames as a cycle, the detection of the first M, n≤M≤T, frame point clouds in this cycle is the initial exploration stage of the algorithm. This stage will explore each action a in the action space k,i , and record each action a k,i The completion delay t k,i ; After the initial exploration phase is completed, the different subtasks are unloaded q k To different nodes b i The task completion delay; After the initial exploration phase is completed, subtask q is performed frame by frame iteratively. k Uninstall action selection, where action a k,i The completion delay t k,i It is estimated as the mean completion time of the same action in all frames before the current frame, and the subtask q is recorded as k The lower bound of the delay of offloading to different nodes is according to Unload the subtasks in order from large to small, that is The largest subtasks are unloaded first. The smallest subtask is uninstalled last; For each subtask q k , select the action a with the highest score W k,i Execution, i.e. subtask q k Uninstall to a k,i The corresponding node b i , specifically: For subtask q k ∈Q, each action a k,i The reward is expressed as in Represents subtask q k The upper bound of the delay when unloading to different nodes, y = U (x) is a simple linear function; At the sth choice, action a k,i The number of times selected is recorded as S k,i (s), action a k,i The uncertainty measure δ of is expressed as follows: The score W is determined by the expected reward and uncertainty measure, namely: Among them, ξ represents the exploration rate; Unload each subtask of the detection task Q to the current best node in turn. After waiting for the subtask to be completed and the parameters to be updated, the detection task of the next frame of point cloud can be carried out. After completing the detection task of T frame point clouds in one cycle, the algorithm is initialized and the next cycle begins.

6. The method according to claim 5, characterized in that ξ is set to λ is the current iteration number.

7. The method according to claim 1, characterized in that Each computing node uses the pre-trained PointPillars model to complete the target detection task.

Images