Trailer motion prediction method and system based on two-dimensional occupancy graph, and medium

Abstract

The invention discloses a towed vehicle motion prediction method and system based on a two-dimensional occupancy graph, and a medium. The prediction method comprises the following steps: generating main features including current state features of each agent; generating interaction features between the agents and interaction features between the agents and a high-precision map through the main features; obtaining the historical state characteristics of each intelligent agent at present; fusing the main feature, the interaction feature between the agents, the interaction feature between the agents and the high-precision map and the historical state feature of the agents to generate a fused feature; and decoding the fused features to obtain a two-dimensional occupancy graph of each agent in a future time period, and predicting the operation of the agents. According to the method, the two-dimensional occupancy map of the object in a period of time in the future can be output, the two-dimensional occupancy map can represent traction vehicles of different types, different shapes and different postures, and the method has the advantages of being high in generalization ability, suitable for respective complex scenes and high in prediction precision.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving, and particularly to a method, system and medium for predicting the movement of a towed vehicle based on a two-dimensional occupancy map. Background Art

[0002] Currently, although the academic end-to-end R & D method has made great progress in the past two years, high-level vehicle autonomous driving still mainly adopts a modular architecture. The core algorithm modules include positioning, perception, prediction, planning and control, etc. The main function of the prediction module is to predict the trajectories of surrounding dynamic traffic participants such as vehicles and pedestrians in the next 5 - 10s based on the current perception results and historical information, so as to provide reference for subsequent planning and decision-making.

[0003] The prediction module is a key module in the autonomous driving system. By predicting the future trajectories and behaviors of surrounding dynamic traffic participants, it can help the autonomous driving vehicle make more reasonable driving decisions. For example, when it is predicted that other vehicles may merge into the current lane, the autonomous driving vehicle can decelerate or change lanes in advance to avoid potential collision risks. This can not only improve the safety of autonomous driving, but also improve the driving efficiency of the vehicle on the premise of ensuring safety.

[0004] A towed vehicle consists of a tractor connected by a hinge to one or more full trailers. Currently, the industry's prediction algorithms basically assume that traffic participants are rigid bodies and predict their future movement trajectory points. These algorithms are not applicable to the scenario of a tractor with a trailer because in the tractor scenario, the tractor and its towed vehicle are different rigid bodies connected together by hinges and their movements are interrelated. The correlation relationship is determined by their connection method, which is completely different from the movements of two independent vehicles.

[0005] Conventional autonomous driving motion prediction models objects as rigid bodies and only anticipates their future trajectory points. In subsequent collision detection, a predefined two-dimensional rectangular box is used to represent their occupied area, which has poor accuracy and cannot represent the tractor and trailer scenario, thus unable to meet the existing requirements. Summary of the Invention

[0006] The purpose of the present invention is to provide a method, system and medium for predicting the movement of a towed vehicle based on a two-dimensional occupancy map to overcome the deficiencies of the prior art. The two-dimensional occupancy map is used to represent the areas occupied by the tractor and its trailer in the future. The shape of the two-dimensional occupancy map is more flexible and applicable to different complex scenarios, and can be directly used for vehicle collision detection and motion planning, which is convenient and accurate to use.

[0007] To achieve the above object, the technical solution adopted by the present invention is: A method for predicting the movement of a towed vehicle based on a two-dimensional occupancy map, comprising the following steps:

[0008] Generate the main features including the current state features of each agent;

[0009] Generate the interaction features between agents and the interaction features between agents and the high-precision map through the main features;

[0010] Obtain the historical state features of each current agent;

[0011] Fuse the main features, the interaction features between agents, the interaction features between agents and the high-precision map, and the historical state features of agents to generate the fused features;

[0012] After decoding the fused features, obtain the two-dimensional occupancy map of each agent in the future time period and then predict the operation of the agent, where the two-dimensional occupancy map is represented as matrix A ixj , and its element A(i,j) represents whether the coordinate (i,j) is occupied. Specifically, 1 - A(i,j) represents the occupancy probability. If the value of 1 - A(i,j) is 0, it means the coordinate point is occupied and other vehicles cannot pass; if the value is 1, it means the coordinate point is not occupied.

[0013] In one embodiment, after obtaining the two-dimensional occupancy map of each agent in the future time period, the main features, the interaction features between agents, the interaction features between agents and the high-precision map, and the historical state features of agents are used to iterate the parameters through the training process to continuously adjust the result of the two-dimensional occupancy map to optimize the LOSS function;

[0014] LOSS = 1 - average IOU(Pi,Qi), where Pi is the predicted two-dimensional occupancy map of the i-th agent in the future, and Qi is the ground truth of the two-dimensional occupancy map of the i-th agent in the future marked manually.

[0015] In one embodiment, the steps to generate the main features including the current state features of each agent are as follows:

[0016] Obtain the perception feature result and the perception vector result in each agent at the current moment;

[0017] Encode the perception feature result and the perception vector result into two perception sub-features respectively;

[0018] Fuse the two perception sub-features to obtain the main features including the current state features of each agent.

[0019] In one embodiment, the perception feature result is the OCC perception result of the current frame. The OCC perception result is an S x L x 3 vector, where S x L is the BEV size and the channels are 3, which are height, whether it is a curb, and the occupancy category respectively.

[0020] In one embodiment, the perception vector results include the position (X, Y, Z), attitude (yaw, pitch, roll), movement speed (vx, vy, vz), and movement acceleration (ax, ay, az) of each agent.

[0021] In one embodiment, the generation steps of the interaction features between the agent and the high-precision map are as follows:

[0022] After the local high-precision map elements are input into the map encoder module, they are converted into local high-precision map features;

[0023] After the local high-precision map features interact with the main features, the interaction features between the agent and the high-precision map are formed.

[0024] In one embodiment, after obtaining the fusion features, the historical feature update module averages the current fusion features with the historical features in a weighted fusion manner to obtain an updated historical feature module.

[0025] A tractor movement prediction system based on a two-dimensional occupancy map includes:

[0026] A main feature generation module, configured to generate main features including the current state features of each agent;

[0027] An interaction feature generation module, configured to generate interaction features between agents and interaction features between agents and the high-precision map through the main features;

[0028] A historical feature acquisition module, configured to acquire the historical state features of each current agent;

[0029] A fusion module, configured to fuse the main features, interaction features between agents, interaction features between agents and the high-precision map, and historical state features of agents to generate fusion features;

[0030] A result output module, configured to decode the fusion features into the movement trajectories of each agent in the future time period based on the two-dimensional occupancy map, where the two-dimensional occupancy map is represented as a matrix A ixj , and its element A(i, j) indicates whether the coordinate (i, j) is occupied. Specifically, 1 - A(i, j) represents the probability of being occupied. When the value of 1 - A(i, j) is 0, it means that the coordinate point is occupied and other vehicles cannot pass; when the value is 1, it means that the coordinate point is not occupied.

[0031] A computer-readable storage medium stores a computer program thereon, and the program is executed by a processor to implement the above-mentioned trailer vehicle movement prediction method based on a two-dimensional occupancy map.

[0032] Due to the application of the above technical solutions, the present invention has the following advantages compared with the prior art:

[0033] The method for predicting the motion of a towed vehicle based on a two-dimensional occupancy map of the present invention first obtains the current main features of each agent and the historical state features of each agent, and then obtains the interaction features between agents and the interaction features between agents and the high-precision map through the main features. Finally, the main features, the interaction features between agents, the interaction features between agents and the high-precision map, and the historical state features of the agents are fused to output the two-dimensional occupancy map of an object for a period of time in the future. The two-dimensional occupancy map can represent tow trucks of different types, shapes, and postures, and has the advantage of strong generalization ability.

[0034] Secondly, after outputting the two-dimensional occupancy maps of each agent at future times and predicting the tractor, it can be directly used for vehicle collision detection and motion planning, which is more convenient and efficient to use, can more accurately predict its future motion trajectory, is more convenient to use, and has higher collision detection accuracy, providing a safer guarantee for autonomous driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings:

[0036] Figure 1 It is a flowchart of the method for predicting the motion of a towed vehicle based on a two-dimensional occupancy map according to an embodiment of the present invention.

[0037] Figure 2 It is a schematic framework diagram of the method for predicting the motion of a towed vehicle based on a two-dimensional occupancy map according to an embodiment of the present invention.

[0038] Figure 3 It is a schematic diagram of the generation process of the main features according to an embodiment of the present invention.

[0039] Figure 4 It is a schematic diagram of a tractor motion prediction system based on a two-dimensional occupancy map according to an embodiment of the present invention.

[0040] Figure 5 It is a comparison diagram of the two-dimensional occupancy map of an empty road and the 2D occupancy of two tractors according to an embodiment of the present invention.

[0041] Figure 6 It is the two-dimensional occupancy map of the tractor at times t1, t2, and t3 according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0043] The present invention provides a method, a system and a medium for predicting the motion of a towed vehicle based on a two-dimensional occupancy map, so as to solve the problem that in the prior art, the industry prediction algorithms basically assume that traffic participants are rigid bodies and predict their future motion trajectory points, but these algorithms cannot be applied to the scenario of a tractor with a trailer because this belongs to two different rigid bodies, resulting in poor prediction efficiency.

[0044] For the sake of easy understanding, the following describes the specific process in the embodiments of the present application. Please refer to Figures 1 to 4 , the method for predicting the motion of a towed vehicle based on a two-dimensional occupancy map in the embodiments of the present application includes the following steps:

[0045] S1. Generate a main feature including the state features of each current agent;

[0046] In step S1, the steps of generating a main feature including the state features of each current agent are as follows: Obtain the perception results in each agent at the current moment. The perception results are divided into a perception feature result and a perception vector result. Encode the perception feature result and the perception vector result into two perception sub-features respectively; finally, fuse the two perception sub-features to obtain a main feature including the state features of each current agent.

[0047] Among them, the perception feature result is the OCC perception result of the current frame. The OCC perception result is a 2000x1000x3 vector. 2000 (S direction) x 1000 (L direction) is the BEV size. Each grid corresponds to 10 cm x 10 cm, and the channels are 3, which are height, whether it is a curb, and the occupied category respectively; the perception vector result includes the position (X, Y, Z), attitude (yaw, pitch, roll), motion speed (vx, vy, vz), and motion acceleration (ax, ay, az) of each agent.

[0048] Refer to Figure 3 , in the process of encoding the perception feature result and the perception vector result into two perception sub-features respectively, a perception sub-feature encoder module and a perception vector feature sub-encoder module are used respectively; the perception sub-feature encoder module is a convolutional deep learning network, and the perception vector feature sub-encoder module is a network based on a transformer, which is a vector feature encoder that converts vector features into fusible vector features.

[0049] In addition, in the process of fusing the two perception sub-features, the perception feature fusion module adopts the attention method. The first perception sub-feature passes through an attention network and is weighted to the second perception sub-feature to obtain the final main feature. The dimension of the main feature is 100x50x128.

[0050] S2. Generate the interaction features between agents and the interaction features between the agent and the high-precision map through the main features;

[0051] In step S2, after the main features are input into the inter-agent interaction module, the inter-agent interaction features are formed. The specific operations are as follows: First, reshape the main features into 2D features, (100x50)x128, then use a 6-layer 3x3 convolutional network, and finally reshape them into 100x50x64 features.

[0052] Secondly, the generation steps of the interaction features between the agent and the high-precision map are as follows: First, the local high-precision map elements are converted into local high-precision map features after being input into the map encoder module. Specifically, the local area range is 200 meters (S direction) x 100 meters (L direction), including elements such as lane lines, curbs, stop lines, crosswalks, etc. The high-precision map elements are represented in a rasterized manner, and the size of each raster is 5cm x 5cm. Therefore, the size of the raster map is 4000x2000xN, where N is the number of feature layers and also the high-precision map element category data. Each layer independently represents the high-precision map element information within the raster. The map encoder module in this embodiment uses a convolutional deep learning network, and after 8 layers of the network, the features are converted into 16x8x256 features.

[0053] Then, the local high-precision map features and the main features interact through the agent and high-precision map interaction module to form the interaction features between the agent and the high-precision map. Specifically, in the agent and high-precision map interaction module, the high-precision map encoded features are regarded as queries, interact with the main features, and then pass through 6 layers of convolutional networks to obtain the interaction features between the agent and the high-precision map. In this embodiment, the dimension of the interaction features is 100x50x24.

[0054] S3. Obtain the historical state features of each current agent;

[0055] In step S3, the agent historical state feature t is the agent historical feature information, which represents the motion state of the agent before the current time t. There are various ways to obtain the historical features of the agent: for example, save all historical moments, or only retain one historical feature, and when the foregoing features come, add them to the historical feature state according to the weight.

[0056] S4. After fusing the main features, the inter-agent interaction features, the interaction features between the agent and the high-precision map, and the agent historical state features, generate the fusion features;

[0057] In step S4, compared with the main features, the fused features contain richer information. There are many ways to fuse features: First, the main features, the inter-agent interaction features, the agent-high definition map interaction features, and the agent historical state features can be directly concatenated for fusion; Second, the main features, the inter-agent interaction features, the agent-high definition map interaction features, and the agent historical state features can also be fused through the attention method. In addition, feature fusion can also be performed simultaneously through the concatenation and attention methods.

[0058] In addition, a hierarchical concatenation method can be adopted, where some features are first fused and then fused with other features. When fusing some features, the above-mentioned concatenation, attention, or the simultaneous use of concatenation and attention fusion methods can be used.

[0059] S5. After decoding the fused features to obtain the two-dimensional occupancy maps of each agent in the future time period, the motion of the agent is predicted.

[0060] In step S5, a prediction module is used for decoding the fused features. The prediction module adopts a decoder module based on a deep learning network and uses an 8-layer convolutional network. The network outputs multiple predicted two-dimensional occupancy maps (2000x1000xM) for each agent, where 2000 (S direction) x 1000 (L direction) is the BEV size, each grid corresponds to 5cm x 5cm, and M is the number of agents to be predicted. The two-dimensional occupancy maps will be passed to the subsequent decision-making and planning module.

[0061] Among them, the decoded two-dimensional occupancy map results are for the future few seconds, generally within 3 to 10 seconds.

[0062] After obtaining the fused features in step S5, there is another step S6. In step S6: The fused features pass through the historical feature update module to update the current historical feature information and prepare for the prediction at the next moment. Here, the historical feature update module averages the current fused features with the historical features using a weighted fusion method to obtain the updated historical feature module. The dimension of the historical feature vector is exactly the same as that of the main feature vector, both being 100 (S direction) x 50 (L direction) x 128.

[0063] In one embodiment, in order to make the two-dimensional occupancy map results output by the prediction more accurate, after obtaining the two-dimensional occupancy maps of each agent in the future time period, the main features, the inter-agent interaction features, the agent-high definition map interaction features, and the agent historical state features are used to iterate the parameters through the training process to continuously adjust the results of the two-dimensional occupancy maps to optimize the LOSS function.

[0064] In this embodiment, LOSS = 1 - average IOU(Pi, Qi), where Pi is the predicted 2D occupancy map of the i-th agent in the future, and Qi is the ground truth of the 2D occupancy map of the i-th agent in the future manually marked. During the calculation of LOSS, first calculate the IOU between the predicted 2D occupancy map and the ground truth, and select the sequence with the largest sum of IOU as the positive sample, and other sequences are negative samples.

[0065] In addition, the number of agents can be automatically adjusted. The current maximum number is 30, and mainly the agents closer to the current host vehicle are selected; when the number of agents is less than 30 or some agents leave the attention area for a long time, the position features are filled with 0.

[0066] After the 2D occupancy map is predicted, the 2D occupancy map is represented as matrix A ixj , and its element A(i, j) indicates whether the coordinate (i, j) is occupied. Specifically, 1 - A(i, j) represents the occupancy probability. For example, a value of 0 means the coordinate point is occupied, and other vehicles cannot pass through; a value of 1 means the coordinate point is not occupied.

[0067] Refer to Figure 5 , in the left figure a, the 2D occupancy map of the road at a moment without traffic participants is shown. The middle blank area is the road area and is not occupied; other cross-sectional areas indicate being occupied and vehicles cannot pass through. The right figure b shows the road occupancy situation when there are two tractors on the road. The two tractors also occupy some areas of the road, and these areas are impassable. Through the tractor operation prediction method based on the 2D occupancy map of the present invention, the 2D occupancy map of the tractor at subsequent times can be estimated, that is, the occupancy probability A of the BEV space of the tractor at multiple subsequent moments t ixj , where t is a subsequent moment.

[0068] Specifically, refer to Figure 6 , from left to right are the 2D occupancy maps of the tractor at times t1, t2, and t3. The prediction of the 2D occupancy map can more accurately represent the occupancy situation of the tractor in the BEV space, can significantly improve the calculation accuracy of vehicle motion planning and collision detection, and this method is not affected by the connection method of the tractor head and the trailer, and has better generalization ability.

[0069] In summary, a method for predicting the motion of a trailer vehicle based on a two-dimensional occupancy map according to the present invention, after obtaining the main features, obtains the interaction features between agents and the interaction features between agents and the high-precision map through the main features, and finally fuses the main features, the interaction features between agents, the interaction features between agents and the high-precision map, and the historical state features of the agents and then decodes them, so as to output the two-dimensional occupancy map (2D Occupancy Map) of an object in a future period of time. The two-dimensional occupancy map is used to represent the area occupied by the tractor and its trailer in the future. The shape of the two-dimensional occupancy map is more flexible, not restricted by rigid bodies, and is also independent of the number of trailers and the towing method. Moreover, the two-dimensional occupancy map can represent tractors with different types, shapes, and postures, and has the advantage of strong generalization ability. Therefore, it is applicable to different complex scenarios, is simpler in form, and is more convenient to apply.

[0070] On the other hand, based on Figure 4 , the present invention also discloses a tractor motion prediction system based on a two-dimensional occupancy map, including: a main feature generation module configured to generate main features including the state features of each current agent; an interaction feature generation module configured to generate the interaction features between agents and the interaction features between agents and the high-precision map through the main features; a historical feature acquisition module configured to acquire the historical state features of each current agent; a fusion module configured to fuse the main features, the interaction features between agents, the interaction features between agents and the high-precision map, and the historical state features of the agents to generate a fusion feature; and a result output module configured to decode the fusion feature into the motion trajectories of each agent in the future time period based on the two-dimensional occupancy map, where the two-dimensional occupancy map is represented as a matrix Ai xj, and its element A(i,j) represents whether the coordinate (i, j) is occupied. Specifically, 1 - A(i,j) represents the probability of being occupied. If the value of 1 - A(i,j) is 0, it means that the coordinate point is occupied and other vehicles cannot pass through; if the value is 1, it means that the coordinate point is not occupied.

[0071] In actual use, for the prediction algorithm of trajectory points, minADE (Minimum Average Displacement Error) is used to evaluate the result error. This patent uses a 2D occupancy map to represent vehicles and mIoU (Mean Intersection over Union) to evaluate the result error. A self-built dataset is used, and the dataset is a closed campus scenario, including trailer vehicles and ordinary vehicles; the dataset includes 8000 samples, of which 80% are used as the training set, 10% are used as the validation set, and 10% are used as the test set. The test results are as follows:

[0072] Algorithm 1smIoU(%) 2smIoU(%) 3smIoU(%) Average This patent 76.28 71.81 67.32 71.80 Vectornet[1] 60.15 53.21 38.67 50.68 M2I[2] 62.37 58.9842 44.30 55.03

[0073] Among them, Vectornet[1] and M2I[2] adopt the trajectory point prediction method. The area is a rectangular box centered on the trajectory point, the direction is the vehicle's forward direction, and the size of the rectangular area is the vehicle size. From the experimental results, the algorithm of this patent can accurately estimate the area occupied by the vehicle in the future, which has great advantages for vehicle motion planning and collision avoidance.

[0074] The experimental results of the next-step collision accuracy also prove this point. Regarding the collision detection accuracy, when comparing the influence of this algorithm and the trajectory point prediction algorithm on motion planning, Planning CR (Collision Rate) is used as the evaluation standard. Planning CR (Collision Rate) refers to the probability or frequency of the path generated by the planning algorithm colliding with obstacles during path planning or motion planning. It is an important indicator for evaluating the safety and reliability of the planning algorithm.

[0075] Algorithm 1sCR(%) 2sCR(%) 3sCR(%) Average This patent 0.01 0.05 0.12 0.06 Vectornet[1] 0.15 1.28 4.69 2.04 M2I[2] 0.12 0.98 3.46 1.52

[0076] From the experimental results, in the semi-trailer scenario, this method can significantly reduce the collision rate compared with the traditional trajectory point prediction algorithm.

[0077] On the other hand, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is used to execute a tractor motion prediction based on a two-dimensional occupancy map provided above. The method includes: generating a main feature including the current state characteristics of each agent; generating an interaction feature between agents and an interaction feature between the agent and the high-precision map through the main feature; obtaining the historical state characteristics of each current agent; fusing the main feature, the interaction feature between agents, the interaction feature between the agent and the high-precision map, and the historical state characteristics of the agent to generate a fusion feature; and decoding the fusion feature to obtain the motion trajectories of each agent in the future time period based on the two-dimensional occupancy map.

[0078] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting the motion of a trailer vehicle based on a two-dimensional occupancy map, characterized in that: The steps include: Generate main features containing the current state features of each intelligent agent; Generate the interaction features between agents and the interaction features between agents and high-precision maps through the main features; Get the historical state characteristics of each current agent; The main features, the interaction features between agents, the interaction features between agents and high-precision maps, and the historical state features of agents are fused to generate fused features; After decoding the fusion features, the two-dimensional occupancy map of each agent in the future time period is obtained to predict the operation of the agent, where the two-dimensional occupancy map is represented by the matrix A ixj , its element A(i,j) indicates whether the coordinate (i, j) is occupied. Specifically, 1-A(i,j) represents the probability of being occupied. A value of 1-A(i,j) of 0 indicates that the coordinate point is occupied, and other vehicles cannot pass through; a value of 1 indicates that the coordinate point is not occupied.

2. The method for predicting the motion of a trailer vehicle based on a two-dimensional occupancy map according to claim 1, characterized in that: After obtaining the two-dimensional occupancy map of each agent in the future time period, the main features, the interaction features between agents, the interaction features between agents and high-precision maps, and the historical state features of agents are used to iterate parameters through the training process to continuously adjust the results of the two-dimensional occupancy map to optimize the LOSS function; LOSS = 1-average IOU(Pi,Qi), where Pi is the predicted two-dimensional occupancy map of the i-th agent in the future, and Qi is the manually labeled true value of the two-dimensional occupancy map of the i-th agent in the future.

3. The method for predicting the motion of a trailer vehicle based on a two-dimensional occupancy map according to claim 1, characterized in that: The steps to generate the main features containing the current state features of each agent are as follows: Obtain the perception feature results and perception vector results of each agent at the current moment; Encode the perceptual feature result and the perceptual vector result into two perceptual sub-features respectively; After fusing the two perception sub-features, the main feature containing the current state features of each intelligent agent is obtained.

4. The method for predicting the movement of a trailer vehicle based on a two-dimensional occupancy map according to claim 3, characterized in that: The perception feature result is the OCC perception result of the current frame. The OCC perception result is an S x L x 3 vector, where SxL is the BEV size and the channel is 3, which are height, curb, and occupancy category.

5. The method for predicting the movement of a trailer vehicle based on a two-dimensional occupancy map according to claim 3, characterized in that: The perception vector results include the position (X, Y, Z), posture (yaw, pitch, roll), movement speed (vx, vy, vz), and movement acceleration (ax, ay, az) of each agent.

6. The method for predicting the movement of a trailer vehicle based on a two-dimensional occupancy map according to claim 1, characterized in that: The steps for generating the interaction features between the intelligent agent and the high-precision map are as follows: The local high-precision map elements are input into the map encoder module and converted into local high-precision map features; The local high-precision map features interact with the main features to form the interactive features of the intelligent agent and the high-precision map.

7. The method for predicting the motion of a trailer vehicle based on a two-dimensional occupancy map according to claim 1, characterized in that: After obtaining the fusion feature, the historical feature update module averages the current fusion feature with the historical feature using a weighted fusion method to obtain an updated historical feature module.

8. A tractor motion prediction system based on a two-dimensional occupancy map, characterized in that: include: A main feature generation module is configured to generate main features including current state features of each intelligent agent; An interactive feature generation module, configured to generate interactive features between agents and interactive features between agents and high-precision maps through main features; A historical feature acquisition module is configured to acquire the historical state features of each current intelligent agent; A fusion module is configured to fuse the main feature, the interaction feature between agents, the interaction feature between agents and high-precision maps, and the historical state feature of the agent to generate a fusion feature; The result output module is configured to decode the fused features into the motion trajectory of each agent in the future time period based on the two-dimensional occupancy map, where the two-dimensional occupancy map is represented by the matrix A ixj , its element A(i,j) indicates whether the coordinate (i, j) is occupied. Specifically, 1-A(i,j) represents the probability of being occupied. A value of 1-A(i,j) of 0 indicates that the coordinate point is occupied, and other vehicles cannot pass through; a value of 1 indicates that the coordinate point is not occupied.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the method for predicting the movement of a towing vehicle based on a two-dimensional occupancy map as described in any one of claims 1 to 8.

Images