Pedestrian intention judgment and trajectory generation method under man-vehicle interaction based on group characteristics

Abstract

The invention discloses a pedestrian intention judgment and track generation method under pedestrian-vehicle interaction based on group features, and belongs to the technical field of automatic driving, and the method comprises the steps: obtaining the pedestrian-vehicle features of a pedestrian-vehicle interaction scene, calculating the space-time proximity and collision time through the pedestrian-vehicle features, and presetting a threshold value to recognize pedestrians and vehicles with interaction; introducing an adjustable distance threshold value as a constraint, and constructing a pedestrian group and a vehicle group which influence pedestrian interaction behaviors; constructing a deep learning network, and extracting respective group characteristics of the pedestrian group and the vehicle group; according to the group features, simulating game interaction behaviors between the pedestrian group and the vehicle group, and fusing the group features and game features between the groups to obtain group fusion features; and according to the group fusion features, pedestrian intention judgment and trajectory generation under human-vehicle interaction are carried out through a deep learning framework. The method can accurately describe the intention of the pedestrian and generate the simulation track, can better assist in planning the driving route, and improves the traffic safety.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to a method for judging pedestrian intentions and generating trajectories in human-vehicle interaction based on group characteristics. Background Art

[0002] With the rapid development of artificial intelligence, sensor technology and vehicle-to-everything (V2X) communications, autonomous driving has become a key innovative technology for intelligent transportation systems. Its core goal is to promote the intelligent and sustainable development of transportation systems by reducing traffic congestion, improving road safety and reducing carbon emissions. In recent years, autonomous driving technology has made significant progress in perception systems, decision-making algorithms and vehicle control, and some autonomous driving functions have been commercialized.

[0003] Although autonomous driving technology has made significant progress, it still faces many challenges in practical applications. First, the safe interaction between autonomous vehicles and pedestrians is the weak link of current technology. Pedestrian behavior is highly uncertain and flexible, and there is a lack of effective interaction mechanism between autonomous vehicles and pedestrians, which makes it difficult to accurately predict pedestrian intentions. Secondly, most existing studies focus on the behavior modeling of individual pedestrians or vehicles, ignoring the impact of group behavior on the interaction process, which makes the generated intention judgment and trajectory prediction not comprehensive enough. In addition, the mechanism model usually has the problem of insufficient accuracy due to the lack of full consideration of complex real-world scenarios. Although the learning model performs better in simulating pedestrian behavior, it still cannot fully cover the complex scenarios in human-vehicle interaction. These shortcomings limit the application of autonomous driving technology in complex traffic environments and reduce the public's trust in autonomous vehicles. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method for judging pedestrian intention and generating trajectory in human-vehicle interaction based on group characteristics, so as to solve the problems existing in the above prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for judging pedestrian intention and generating trajectory in human-vehicle interaction based on group characteristics, comprising:

[0006] Obtaining human-vehicle features in a human-vehicle interaction scene, calculating spatiotemporal proximity and collision time based on the human-vehicle features, and identifying pedestrians and vehicles that are interacting based on preset thresholds of spatiotemporal proximity and collision time;

[0007] An adjustable distance threshold is introduced as a constraint to construct pedestrian groups and vehicle groups that affect pedestrian interaction behaviors; a deep learning network is constructed to extract the group characteristics of pedestrian groups and vehicle groups;

[0008] According to the group characteristics, the game interaction behavior between the pedestrian group and the vehicle group is simulated, and the group characteristics are integrated with the game characteristics between the groups to obtain the group integration characteristics;

[0009] According to the group fusion characteristics, pedestrian intention judgment and trajectory generation under human-vehicle interaction are performed through a deep learning framework.

[0010] Preferably, identifying pedestrians and vehicles that are interacting includes:

[0011] Obtaining human-vehicle features of a human-vehicle interaction scene; the human-vehicle features include pedestrian position features, vehicle position features, pedestrian speed features, and vehicle speed features;

[0012] The spatiotemporal proximity is calculated by using the human-vehicle characteristics and recorded in a symmetrical spatiotemporal proximity matrix;

[0013] The collision time is calculated by using the human-vehicle characteristics and recorded in a symmetrical collision relative time matrix;

[0014] Based on the spatiotemporal proximity matrix and the collision relative time matrix, adjustable thresholds are respectively set, and pedestrians and vehicles that interact are identified according to the thresholds.

[0015] Preferably, extracting the group characteristics of the pedestrian group and the vehicle group includes:

[0016] An adjustable distance threshold is introduced as a constraint to filter out surrounding pedestrians and surrounding vehicles that affect the interactive pedestrian behavior;

[0017] Based on the surrounding pedestrians and the surrounding vehicles, introducing interactive pedestrians to construct pedestrian groups and vehicle groups;

[0018] A scene-adaptive deep learning network is constructed respectively, and group features of the pedestrian group and the vehicle group are extracted through the deep learning network.

[0019] Preferably, the group characteristics include group internal state relations and group position relations;

[0020] The calculation formula of the internal state relationship of the group is:

[0021] ;

[0022] in, and are the state characteristics of each subject within the pedestrian group or vehicle group, and are learnable parameters, is the dimension of state characteristics;

[0023] The calculation formula of the group position relationship is:

[0024] ;

[0025] in, is the spatial distance between the entities within the pedestrian group or vehicle group; is an adjustable space threshold; is the indicator function.

[0026] Preferably, obtaining the group fusion feature includes:

[0027] According to the group characteristics, a game strategy set of pedestrian groups and vehicle groups is constructed; through the game strategy set, the game utility of the pedestrian group and the vehicle group is established;

[0028] Based on the game utility, the game interaction behaviors of pedestrian groups and vehicle groups under human-vehicle interaction are simulated to obtain the expected utility of pedestrian groups and vehicle groups;

[0029] Based on the expected utility, the group features of the pedestrian group and the vehicle group are fused to obtain a group fusion feature.

[0030] Preferably, the game utility includes direct risk of collision, indirect risk, avoidance loss and group utility; the indirect risk includes minimum future relative distance, minimum future relative time, and greater participant conflict speed.

[0031] Preferably, the game utility of the pedestrian group and the vehicle group is expressed as:

[0032] ;

[0033] in, For pedestrians or vehicles The utility under the strategy combination, For pedestrians or vehicles The minimum future relative distance under the strategy combination, For pedestrians or vehicles The minimum future relative time under the strategy combination, For pedestrians or vehicles The maximum participant conflict speed under the strategy combination.

[0034] Preferably, the process of pedestrian intention judgment and trajectory generation under human-vehicle interaction through a deep learning framework includes:

[0035] Taking the group fusion features as input, a behavior intention feature extraction module under a deep learning framework is constructed, and the behavior intention features in the human-vehicle interaction scene are extracted;

[0036] Performing shaping and scaling on the behavior intention feature to obtain a reward function for the interacting pedestrian under the human-vehicle interaction;

[0037] Using the reward function as a criterion, the behavior strategy of the interactive pedestrian under the influence of the pedestrian group and the vehicle group is obtained;

[0038] The behavior intention feature extraction module and the behavior strategy are iteratively updated to judge pedestrian intentions in heterogeneous interaction scenarios and generate pedestrian trajectories.

[0039] In a second aspect, the present invention further discloses a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method described in the first aspect when executed by a processor.

[0040] In a third aspect, the present invention further discloses a computer program product, comprising a computer program, which implements the steps of the method described in the first aspect when executed by a processor.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] The present invention provides a method for judging the intention of pedestrians and generating trajectories in human-vehicle interaction based on group characteristics. First, the human-vehicle characteristics of the human-vehicle interaction scene are obtained, and the spatiotemporal proximity and collision time are calculated through the human-vehicle characteristics. The pedestrians and vehicles in interaction are identified through preset thresholds of the spatiotemporal proximity and collision time; secondly, an adjustable distance threshold is introduced as a constraint to construct a pedestrian group and a vehicle group that affect the pedestrian interaction behavior; a deep learning network is constructed, and the group characteristics of each pedestrian group and vehicle group are extracted; further, according to the group characteristics, the game interaction behavior between the pedestrian group and the vehicle group is simulated, and the group characteristics are integrated with the game characteristics between the groups to obtain the group fusion characteristics; finally, according to the group fusion characteristics, the pedestrian intention judgment and trajectory generation in human-vehicle interaction are performed through a deep learning framework.

[0043] The present invention accurately identifies pedestrians and vehicles interacting in human-vehicle interaction scenarios through threshold indicators of spatiotemporal proximity and collision time; by constructing pedestrian groups and vehicle groups, a scene-adaptive deep learning framework is established to accurately extract the group characteristics of pedestrian groups and vehicle groups. Subsequently, the game interaction behavior between pedestrian groups and vehicle groups is introduced, and the group characteristics of each group are deeply integrated. Finally, the deep learning framework is used to capture the behavioral intentions and optimal strategies of pedestrians in human-vehicle interaction in the fused features, so as to accurately judge the behavioral intentions and generate high-precision pedestrian trajectories.

[0044] The present invention can effectively handle the influence of surrounding pedestrians and vehicles on pedestrian behavior in human-vehicle interaction, accurately describe pedestrian intentions and generate simulated trajectories, and can better assist in planning driving routes and improve traffic safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0046] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of indirect risk of human-vehicle interaction according to an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of reward function analysis for interactive pedestrians according to an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of interactive pedestrian trajectories generated according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0052] Embodiment 1

[0053] like Figure 1 As shown, this embodiment provides a method for judging pedestrian intention and generating trajectory in human-vehicle interaction based on group characteristics, including:

[0054] S1. Extracting human-vehicle features from a human-vehicle interaction scene, calculating the spatiotemporal proximity and collision time through the human-vehicle features, and identifying pedestrians and vehicles interacting through preset thresholds of the spatiotemporal proximity and collision time; wherein the human-vehicle features include pedestrian position features, vehicle position features, pedestrian speed features, and vehicle speed features.

[0055] Specifically, S1 includes:

[0056] S11. Extract pedestrian position features, vehicle position features, pedestrian speed features, vehicle speed features and other features from the human-vehicle interaction scene, calculate paired spatiotemporal proximity indicators based on the obtained features of pedestrian i and vehicle j, and record the paired spatiotemporal proximity in a symmetric proximity matrix.

[0057] The temporal-spatial proximity index may be the Euclidean distance maintained between pedestrians and vehicles over a period of time.

[0058] S12. Calculate the paired relative collision times based on the obtained features of the pedestrian i and the vehicle j, and record the paired relative collision times in a symmetrical paired relative collision time matrix.

[0059] The relative time of collision can be the absolute time difference between pedestrian i and vehicle j passing through the intersection of their trajectories (if any), which is defined as follows:

[0060] ;

[0061] In the formula, is the collision relative time, is the moment when pedestrian i passes the intersection point with vehicle j’s trajectory, is the moment when vehicle j passes the intersection with the trajectory of pedestrian i.

[0062] S13. Based on the spatiotemporal proximity matrix and the collision relative time matrix, an adjustable threshold is set to identify pairs of pedestrians and vehicles that interact, and the interaction categories and numbers of the pedestrians and vehicles that interact are recorded.

[0063] The adjustable threshold may include two aspects. On the one hand, it is a threshold for the temporal and spatial proximity set in S11, that is, the spatial distance (e.g., 10 meters) that pedestrian i and vehicle j should maintain within a certain period of time (e.g., 2 seconds). The spatial distance can be determined and calculated based on the braking performance of the vehicle and the speed limit of the vehicle during operation. On the other hand, it is a threshold for the relative time of collision, which can be considered to have a lower The scenario is a relatively dangerous scenario. In actual application, 3 seconds can be taken as the threshold.

[0064] S2. Introduce an adjustable distance threshold as a constraint, construct pedestrian groups and vehicle groups that affect pedestrian interaction behaviors, build a deep learning network, and extract the group characteristics of the pedestrian group and vehicle group.

[0065] Specifically, S2 includes:

[0066] S21, calculating the pairwise distance between pedestrian i and the interacting pedestrians in the scene, introducing an adjustable distance threshold under pedestrian interaction, screening surrounding pedestrians that have an impact on the behavior of the interacting pedestrians, and recording them;

[0067] The pairwise distance may be the Euclidean distance between pedestrian i and the interacting pedestrian, and the adjustable distance threshold may be the middle value between the intimate distance and the social distance in social relationship theory. In practical application, 4 m may be taken.

[0068] S22, calculating the paired distance between vehicle i and the interacting pedestrian in the scene, introducing an adjustable distance threshold under human-vehicle interaction, screening surrounding vehicles that affect the behavior of the interacting pedestrian, and recording them;

[0069] The paired distance may be the Euclidean distance between vehicle i and the interacting pedestrian, and the adjustable distance threshold may be determined based on the speed limit of the road section, the traffic flow, and the braking performance of the vehicle. In actual application, 10 meters may be taken.

[0070] S23. As an innovative implementation method, based on the surrounding pedestrians and the surrounding vehicles, interactive pedestrians are introduced to construct pedestrian groups and vehicle groups, and scene-adaptive deep learning network frameworks are respectively constructed to extract group features of pedestrian groups and vehicle groups.

[0071] The pedestrian group should include interactive pedestrians in human-vehicle interaction and all surrounding pedestrians that meet the pedestrian interaction adjustable distance threshold determined by the judgment;

[0072] The vehicle group should include interactive pedestrians and interactive vehicles under human-vehicle interaction, and all surrounding vehicles that meet the adjustable distance threshold for human-vehicle interaction as determined.

[0073] Due to the differences in dynamic characteristics, appearance, etc., there should be different interaction modes within the pedestrian group and the vehicle group. Based on this, this embodiment constructs a deep learning network framework for the pedestrian group and the vehicle group to extract the group characteristics of the pedestrian group and the vehicle group in the human-vehicle interaction scene.

[0074] Specifically, for each group, the internal state relationship of the group is first calculated based on the extracted state features such as contour, speed, acceleration, orientation, and position features such as position. Relationship with group position .

[0075] Internal state relations of a group The calculation can be performed by dot product, embedded dot product, relationship graph, etc. In this embodiment, the embedded dot product is used to calculate the internal state relationship of the group, as follows:

[0076] ;

[0077] In the formula, and are the state characteristics of each subject within the pedestrian group or vehicle group, and are learnable parameters, is the dimension of the state feature.

[0078] Position relationship within the group The calculation can be performed by means of position mask, embedded dot product, dot product, etc. In this embodiment, the position mask is used to calculate the position relationship within the group, which is as follows:

[0079] ;

[0080] In the formula, is the spatial distance between the entities in the pedestrian group or vehicle group. In practical applications, the Euclidean distance can be used; is an adjustable space threshold; is an indicator function, which takes the value 1 if the condition is met, and 0 otherwise.

[0081] Through the state features and the position features, the internal influence relationship of the pedestrian group or vehicle group can be further calculated by dot product, exponential function dot product, embedded dot product, etc. This embodiment uses the form of exponential function dot product combined with softmax function for explanation, which is as follows:

[0082] ;

[0083] Based on the internal influence relationship of the pedestrian group or vehicle group, graph convolution, multi-layer perceptron, convolution, etc. can be further used to extract features. This embodiment uses graph convolution to aggregate its internal features, as follows:

[0084] ;

[0085] In the formula, is the characteristic representation of the pedestrian or vehicle group at the lth layer, is the learnable parameter matrix.

[0086] S3. According to the group characteristics, simulate the game interaction behavior between the pedestrian group and the vehicle group, and integrate the group characteristics with the game characteristics between the groups to obtain group fusion characteristics.

[0087] Specifically, S3 includes:

[0088] S31. As an innovative implementation method, a game strategy set of pedestrian groups and vehicle groups is constructed, and based on factors such as speed characteristics, group size, etc. of the pedestrian groups and vehicle groups, the game utility of all strategy combinations of the pedestrian groups and vehicle groups is established.

[0089] The game strategy set may include deceleration, deflection, acceleration, etc. In this embodiment, passing (maintaining the original state) and giving way (deflection and deceleration) are adopted as the strategy sets for pedestrian groups and vehicle groups.

[0090] The game utility may include direct risk of collision, indirect risk, avoidance loss and group utility. In this embodiment, the indirect risks of the minimum future relative distance, the minimum future relative time and the maximum participant conflict speed are selected to describe the game interaction behavior of the pedestrian group and the vehicle group. The schematic diagram of the three indirect risks is as follows: Figure 2 shown.

[0091] Minimum future relative distance evaluates the distance of the other participant from the potential conflict point when the pedestrian or vehicle arrives at the potential conflict point. Minimum future relative time evaluates the remaining time for the other participant to reach the potential conflict point when the pedestrian or vehicle arrives at the potential conflict point. Larger participant conflict speed evaluates the current speed of the vehicle when the pedestrian or vehicle arrives at the potential conflict point.

[0092] When pedestrian groups and vehicle groups adopt different strategies, the speed, direction, and location of potential conflict points will change. Therefore, under different strategy combinations of pedestrian groups and vehicle groups, the game utility will change. The game utility of pedestrian groups and vehicle groups can be expressed as follows:

[0093] ;

[0094] In the formula, For pedestrians or vehicles The utility under the strategy combination, For pedestrians or vehicles The minimum future relative distance under the strategy combination, For pedestrians or vehicles The minimum future relative time under the strategy combination, For pedestrians or vehicles The maximum participant conflict speed under the strategy combination.

[0095] S32. Based on the game utility, simulate the game interaction behaviors of the pedestrian group and the vehicle group in the human-vehicle interaction to obtain the expected utility of the pedestrian group and the vehicle group.

[0096] Based on the game utility, the expected utility of pedestrians and vehicles after adopting a certain strategy is further calculated as follows:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] In the formula, and They are the expected utilities of the pedestrian group adopting the passing and yielding game strategies, and the same is true for the vehicle group. and The probability of taking a passing strategy for pedestrian groups and vehicle groups.

[0102] According to the expected utility of the pedestrian group and the vehicle group adopting the passing or yielding strategy, the expected utility of the pedestrian and vehicle group in the pedestrian-vehicle interaction downward movement in this state is calculated as follows:

[0103] ;

[0104] ;

[0105] In the formula, and The expected utility of pedestrian and vehicle groups under human-vehicle interaction.

[0106] According to the expected utility, the behavior of pedestrians and vehicles can be simulated by Nash game, evolutionary game, master-slave game, etc. In this embodiment, the behavior of pedestrians and vehicles is simulated by evolutionary game, and the replication dynamic equation of evolutionary game is as follows:

[0107] ;

[0108] ;

[0109] In the formula, and They are the replicated dynamic equations for pedestrian groups and vehicle groups respectively, which can be combined with the strategy distribution and expected utility of the current state of pedestrian groups and vehicle groups to simulate the game behavior of pedestrian groups and vehicle groups.

[0110] S33. Based on the expected utility, the group characteristics of the pedestrian group and the vehicle group are fused to obtain a group fusion feature.

[0111] After the evolutionary game model of the pedestrian group and the vehicle group is calibrated by the genetic algorithm, the corresponding utility equations are:

[0112] ;

[0113] ;

[0114] According to the utility of the pedestrian group and the vehicle group under each strategy combination, after the expected utility is calculated, it is fused with the group characteristics of the pedestrian group and the vehicle group. This can be done by dot product, cross product, embedded dot product, etc. In this embodiment, the dot product form is used, which is as follows:

[0115] ;

[0116] Where F is the group fusion characteristic.

[0117] S4. Based on the group fusion characteristics, pedestrian intention judgment and trajectory generation under human-vehicle interaction are performed through a deep learning framework.

[0118] Specifically, S4 includes:

[0119] S41. As an innovative implementation method, the group fusion features are used as input to construct a deep learning framework behavior intention feature extraction module to extract the behavior intention features in the human-vehicle interaction scenario.

[0120] The deep learning framework may adopt a network architecture such as a multi-layer perceptron, a convolutional network, etc. In this embodiment, a multi-layer perceptron is used to extract the group fusion features to obtain the behavioral intention features of the interacting pedestrians, as follows:

[0121] ;

[0122] In the formula, It is the reward value of the interacting pedestrian after being processed by the deep learning framework, which can be used to preliminarily judge the pedestrian's intention.

[0123] S42. As an innovative implementation method, a deep learning algorithm module is established to perform shaping and scaling on the behavior intention features to obtain a reward function for interactive pedestrians in human-vehicle interaction.

[0124] Based on the reward value of the interactive pedestrian processed by the deep learning framework, the deep learning network framework can be further used to perform shaping calculations. Here, the generative adversarial learning framework can be used to calculate it and obtain the reward function of the interactive pedestrian under human-vehicle interaction. The details are as follows:

[0125] ;

[0126] ;

[0127] ;

[0128] In the formula, An integer term to help adjust the reward value, is the discount factor, To determine whether the current trajectory is a real trajectory, is the probability of taking action a in state s during training, is the reward function after shaping and scaling.

[0129] S43. Using the reward function as a criterion, a behavior strategy of the interactive pedestrian under the influence of the pedestrian group and the vehicle group is obtained.

[0130] Taking the reward function as the evaluation criterion, reinforcement learning algorithms such as Q-network, deep Q-network, and soft actor-critic can be used to maximize the reward function to obtain the optimal behavior strategy of interactive pedestrians under the influence of pedestrian groups and vehicle groups in human-vehicle interaction.

[0131] S44, iteratively updating the behavior intention feature extraction module and the behavior strategy, performing pedestrian intention judgment in heterogeneous interaction scenarios and generating pedestrian trajectories.

[0132] The pedestrian intention judgment and trajectory generation methods are both dynamic processes, focusing on judging the behavior intention according to the current time t and generating the pedestrian trajectory at time t+1.

[0133] The reward function and the behavior strategy are a feedback optimization process. After the algorithm converges, the reward function analysis and pedestrian trajectory generation can be performed, respectively as follows: Figure 3 and Figure 4 shown.

[0134] Beneficial effects of this embodiment:

[0135] This embodiment uses adjustable spatiotemporal proximity and collision time threshold indicators to accurately identify pedestrians and vehicles interacting in human-vehicle interaction scenes, and further combines heterogeneous adjustable spatial distance indicators to identify surrounding pedestrians and vehicles that affect interactive pedestrians, and thereby constructs pedestrian groups and vehicle groups, and establishes a scene-adaptive deep learning framework, thereby accurately extracting group characteristics between individuals within pedestrian groups and vehicle groups. Subsequently, the game interaction behavior between pedestrian groups and vehicle groups is introduced, and group characteristics and game characteristics between groups are deeply integrated. Finally, the deep learning framework is used to capture the behavioral intentions and optimal strategies of pedestrians in human-vehicle interaction in the fused features, thereby accurately judging behavioral intentions and generating high-precision pedestrian trajectories.

[0136] Embodiment 2

[0137] This embodiment further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the first embodiment are implemented.

[0138] Embodiment 3

[0139] This embodiment also discloses a computer program product, including a computer program, which implements the steps of the method described in the first embodiment when executed by a processor.

[0140] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for judging pedestrian intention and generating trajectory in human-vehicle interaction based on group characteristics, characterized in that: The following steps are involved: Obtaining human-vehicle features in a human-vehicle interaction scene, calculating spatiotemporal proximity and collision time based on the human-vehicle features, and identifying pedestrians and vehicles that are interacting based on preset thresholds of spatiotemporal proximity and collision time; An adjustable distance threshold is introduced as a constraint to construct pedestrian groups and vehicle groups that affect pedestrian interaction behaviors; a deep learning network is constructed to extract the group characteristics of pedestrian groups and vehicle groups; According to the group characteristics, the game interaction behavior between the pedestrian group and the vehicle group is simulated, and the group characteristics are integrated with the game characteristics between the groups to obtain the group integration characteristics; According to the group fusion characteristics, pedestrian intention judgment and trajectory generation under human-vehicle interaction are performed through a deep learning framework.

2. The method according to claim 1, characterized in that Identifying pedestrians and vehicles that are interacting includes: Obtaining human-vehicle features of a human-vehicle interaction scene; the human-vehicle features include pedestrian position features, vehicle position features, pedestrian speed features, and vehicle speed features; The spatiotemporal proximity is calculated by using the human-vehicle characteristics and recorded in a symmetrical spatiotemporal proximity matrix; The collision time is calculated by using the human-vehicle characteristics and recorded in a symmetrical collision relative time matrix; Based on the spatiotemporal proximity matrix and the collision relative time matrix, adjustable thresholds are respectively set, and pedestrians and vehicles that interact are identified according to the thresholds.

3. The method according to claim 1, characterized in that Extracting the group characteristics of pedestrian groups and vehicle groups includes: An adjustable distance threshold is introduced as a constraint to filter out surrounding pedestrians and surrounding vehicles that affect the interactive pedestrian behavior; Based on the surrounding pedestrians and the surrounding vehicles, introducing interactive pedestrians to construct pedestrian groups and vehicle groups; A scene-adaptive deep learning network is constructed respectively, and group features of the pedestrian group and the vehicle group are extracted through the deep learning network.

4. The method according to claim 3, characterized in that The group characteristics include the internal state relationship of the group and the position relationship of the group; The calculation formula of the internal state relationship of the group is: ； in, and are the state characteristics of each subject within the pedestrian group or vehicle group, and are learnable parameters, is the dimension of state characteristics; The calculation formula of the group position relationship is: ； in, is the spatial distance between the entities within the pedestrian group or vehicle group; is an adjustable space threshold; is the indicator function.

5. The method according to claim 1, characterized in that The obtained group fusion features include: According to the group characteristics, a game strategy set of pedestrian groups and vehicle groups is constructed; through the game strategy set, the game utility of the pedestrian group and the vehicle group is established; Based on the game utility, the game interaction behaviors of pedestrian groups and vehicle groups under human-vehicle interaction are simulated to obtain the expected utility of pedestrian groups and vehicle groups; Based on the expected utility, the group features of the pedestrian group and the vehicle group are fused to obtain a group fusion feature.

6. The method according to claim 5, characterized in that Game utility includes direct risk of collision, indirect risk, avoidance loss and group utility; the indirect risk includes minimum future relative distance, minimum future relative time and maximum participant conflict speed.

7. The method according to claim 6, characterized in that The game utility of pedestrian groups and vehicle groups considering indirect risks is expressed as: ； in, For pedestrians or vehicles The utility under the strategy combination, For pedestrians or vehicles The minimum future relative distance under the strategy combination, For pedestrians or vehicles The minimum future relative time under the strategy combination, For pedestrians or vehicles The maximum participant conflict speed under the strategy combination.

8. The method according to claim 1, characterized in that The process of pedestrian intention judgment and trajectory generation in human-vehicle interaction through a deep learning framework includes: Taking the group fusion features as input, a behavior intention feature extraction module under a deep learning framework is constructed, and the behavior intention features in the human-vehicle interaction scene are extracted; Performing shaping and scaling on the behavior intention feature to obtain a reward function for the interacting pedestrian under the human-vehicle interaction; Using the reward function as a criterion, the behavior strategy of the interactive pedestrian under the influence of the pedestrian group and the vehicle group is obtained; The behavior intention feature extraction module and the behavior strategy are iteratively updated to judge pedestrian intentions in heterogeneous interaction scenarios and generate pedestrian trajectories.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images