Vehicle control method and vehicle

Abstract

The present disclosure provides a vehicle control method and a vehicle. The method comprises: for at least one agent, based on a current observation period historical trajectory, combining cross-period reference information of a previous observation period, predicting a trajectory of a future period corresponding to the current observation period, and obtaining a fusion prediction trajectory corresponding to the current observation period; if the current observation period does not satisfy a prediction terminal step condition, generating cross-period reference information of the current observation period, and moving the observation period by a preset step to obtain a next observation period; if the current observation period satisfies the prediction terminal step condition, determining a trajectory prediction result of the at least one agent as the fusion prediction trajectory corresponding to the current observation period; and determining a driving strategy of the vehicle and controlling the vehicle to drive according to the trajectory prediction result of the at least one agent. The present disclosure can improve the reliability and safety of vehicle control.

Description

Technical Field

[0001] This disclosure relates to the fields of smart mining, autonomous driving, and vehicle technology, and in particular to a vehicle control method and a vehicle. Background Technology

[0002] When controlling a vehicle, by predicting the future trajectories of other intelligent agents (such as vehicles and pedestrians) in the traffic scene, changes in the surrounding traffic situation can be perceived in advance, improving the rationality, foresight, and safety of the vehicle's driving decisions.

[0003] However, existing trajectory prediction methods are not accurate enough, which in turn leads to a decrease in the reliability and safety of vehicle control. Summary of the Invention

[0004] This disclosure provides a vehicle control method and a vehicle to address the shortcomings in reliability and safety of existing vehicle control methods.

[0005] In view of the above problems, a first aspect provides a vehicle control method, the method comprising:

[0006] For at least one agent, based on the historical trajectory of the current observation period and combined with the cross-segment reference information of the previous observation period, the trajectory of the future period corresponding to the current observation period is predicted to obtain the fused predicted trajectory corresponding to the current observation period.

[0007] If the current observation period does not meet the prediction end-step condition, cross-segment reference information for the current observation period is generated, and the observation period is shifted to the next observation period according to the preset step size.

[0008] If the current observation period meets the prediction final step condition, the trajectory prediction result of the at least one agent is determined as the fusion prediction trajectory corresponding to the current observation period.

[0009] Based on the trajectory prediction results of at least one intelligent agent, the vehicle's driving strategy is determined and the vehicle's driving is controlled.

[0010] In conjunction with the first aspect, in one possible implementation, the step of predicting the trajectory of future periods corresponding to the current observation period based on the historical trajectory of the current observation period and in conjunction with the cross-segment reference information of the previous observation period to obtain the fused predicted trajectory corresponding to the current observation period includes:

[0011] Based on the historical trajectory of the current observation period, the trajectory of the future period corresponding to the current observation period is predicted to obtain the basic predicted trajectory corresponding to the current observation period.

[0012] Based on the basic prediction trajectory corresponding to the previous observation period, the basic prediction trajectory corresponding to the current observation period is corrected to obtain the fused prediction trajectory corresponding to the current observation period.

[0013] In conjunction with the first aspect, in one possible implementation, the step of correcting the basic prediction trajectory corresponding to the current observation period based on the basic prediction trajectory corresponding to the previous observation period to obtain the fused prediction trajectory corresponding to the current observation period includes:

[0014] The decoded features of the basic predicted trajectory corresponding to the current observation period are used as query-side features, and the decoded features of the basic predicted trajectory corresponding to the previous observation period are used as reference-side features. Cross-segment attention calculation is performed to obtain feature information for correcting the predicted trajectory.

[0015] Based on the feature information used to correct the predicted trajectory, the basic predicted trajectory corresponding to the current observation period is corrected to obtain the fused predicted trajectory corresponding to the current observation period.

[0016] In conjunction with the first aspect, in one possible implementation, the step of correcting the basic predicted trajectory corresponding to the current observation period based on the feature information used to correct the predicted trajectory, to obtain the fused predicted trajectory corresponding to the current observation period, includes:

[0017] The feature information used to correct the predicted trajectory is regressed into the correction amount of each position in the basic predicted trajectory corresponding to the current observation period through the feedforward network.

[0018] For each position in the basic predicted trajectory corresponding to the current observation period, the correction amount of that position is superimposed on the original coordinates of that position to obtain the fused predicted trajectory corresponding to the current observation period.

[0019] In conjunction with the first aspect, in one possible implementation, after predicting the trajectory of future periods corresponding to the current observation period based on the historical trajectory of the current observation period to obtain the basic predicted trajectory corresponding to the current observation period, the method further includes:

[0020] Using the actual trajectory of the future time period as a supervision signal, the model parameters used to generate the basic predicted trajectory are optimized to minimize the error between the basic predicted trajectory and the actual trajectory.

[0021] In conjunction with the first aspect, in one possible implementation, the step of predicting the trajectory of future periods corresponding to the current observation period based on the historical trajectory of the current observation period and in conjunction with the cross-segment reference information of the previous observation period to obtain the fused predicted trajectory corresponding to the current observation period includes:

[0022] Based on the historical trajectory of the current observation period, obtain the basic feature information of the current observation period;

[0023] Based on the fused feature information of the previous observation period, the basic feature information of the current observation period is corrected to obtain the fused feature information of the current observation period.

[0024] Based on the fusion feature information of the current observation period, the trajectory of the future period corresponding to the current observation period is predicted, and the fusion prediction trajectory corresponding to the current observation period is obtained.

[0025] In conjunction with the first aspect, in one possible implementation, the step of correcting the basic feature information of the current observation period based on the fused feature information of the previous observation period to obtain the fused feature information of the current observation period includes:

[0026] Using the basic feature information of the current observation period as the query-side feature and the fusion feature information of the previous observation period as the reference-side feature, cross-segment attention calculation is performed to obtain the fusion feature information of the current observation period.

[0027] In conjunction with the first aspect, in one possible implementation, after predicting the trajectory of the future time period corresponding to the current observation period based on the historical trajectory of the current observation period and in combination with the cross-segment reference information of the previous observation period to obtain the fused predicted trajectory corresponding to the current observation period, the method further includes:

[0028] Using the real trajectory corresponding to the future time period as a supervision signal, the model parameters used to generate the fused predicted trajectory are optimized to minimize the error between the fused predicted trajectory and the real trajectory.

[0029] In conjunction with the first aspect, in one possible implementation, the current observation period satisfies the prediction final step condition, including at least one of the following:

[0030] The end of the current observation period coincides with the current time.

[0031] The current observation period has reached the preset number of prediction steps;

[0032] The fusion prediction trajectory corresponding to the current observation period does not overlap with the historical trajectory in time.

[0033] Secondly, a vehicle is provided, comprising:

[0034] Memory, which stores computer programs;

[0035] A processor for running the program, wherein the program, when running, executes the vehicle control method as described in the first aspect, or in any possible implementation of the first aspect.

[0036] Thirdly, a computer device is provided, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and the machine-readable instructions, when executed by the processor, perform steps of the vehicle control method as described in the first aspect, or in combination with any possible embodiment of the first aspect.

[0037] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, performs the steps of the vehicle control method as described in the first aspect, or in conjunction with any possible embodiment of the first aspect.

[0038] The beneficial effects of the embodiments disclosed herein include:

[0039] This disclosure provides a vehicle control method and a vehicle. By progressively iteratively predicting the trajectory of future time periods through multiple observation periods, and by fusing cross-segment reference information from the previous observation period to generate a corresponding fused prediction trajectory during the trajectory prediction process of each observation period, the method achieves temporal correlation and information fusion between the observation periods. This avoids the error accumulation problem caused by the traditional single-step prediction of the complete future trajectory and enhances the adaptive response capability to dynamic changes in the trajectory of the intelligent agent, thereby improving the accuracy of the trajectory prediction results and enhancing the reliability and safety of vehicle driving decision-making and control. Attached Figure Description

[0040] Figure 1 A flowchart of a vehicle control method provided in an embodiment of this disclosure;

[0041] Figure 2 A flowchart of a trajectory prediction method provided in this embodiment of the disclosure;

[0042] Figure 3 A schematic diagram of trajectory prediction results of a benchmark model provided in an embodiment of this disclosure;

[0043] Figure 4 A schematic diagram of trajectory prediction results for an improved model provided in an embodiment of this disclosure;

[0044] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0045] This disclosure provides a vehicle control method and a vehicle. Preferred embodiments of this disclosure are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of this disclosure. Furthermore, the embodiments and features described herein can be combined with each other unless otherwise specified.

[0046] This disclosure provides a vehicle control method. Figure 1 A flowchart of a vehicle control method provided in this disclosure embodiment is shown below. Figure 1 As shown, the method includes:

[0047] S101. For at least one intelligent agent, based on the historical trajectory of the current observation period and combined with the cross-segment reference information of the previous observation period, predict the trajectory of the future period corresponding to the current observation period to obtain the fusion prediction trajectory corresponding to the current observation period.

[0048] S102. If the current observation period does not meet the prediction end-step condition, generate cross-segment reference information for the current observation period, and shift the observation period backward according to the preset step size to obtain the next observation period.

[0049] S103. If the current observation period meets the prediction final step condition, determine the trajectory prediction result of at least one agent as the fusion prediction trajectory corresponding to the current observation period.

[0050] S104. Based on the trajectory prediction results of at least one agent, determine the vehicle's driving strategy and control the vehicle's driving.

[0051] In this embodiment, trajectory prediction is performed on an intelligent agent, which is an object with motion attributes in a traffic scene, including at least one of motor vehicles, non-motor vehicles, and pedestrians. The trajectory prediction result of the intelligent agent serves as the basis for determining the vehicle's driving strategy. Based on the future movement trend of the intelligent agent, changes in the traffic scene can be predicted, and the vehicle's driving behavior can be rationally planned to avoid driving conflicts with other intelligent agents. For example, when a pedestrian is predicted to cross the road, the vehicle can adopt driving strategies such as slowing down to avoid them or stopping to wait.

[0052] In some embodiments, trajectory prediction can be performed on the agent closest to the vehicle, or it can be performed on all agents within a preset driving range (such as agents on the vehicle's driving path).

[0053] In the field of data-driven trajectory prediction, mainstream trajectory prediction models generally adopt a direct prediction strategy with a fixed time step during the inference process. Specifically, the model usually takes historical trajectories of a certain time length as input and outputs the complete future trajectory within the target time range at once. Although this "single-step inference - full-step output" paradigm is simple to implement and has a relatively fast inference speed, since the model generates the future trajectory at once, errors tend to accumulate over time, especially in long-term range predictions, which significantly reduces the accuracy of the trajectory prediction results and the reliability and safety of vehicle control.

[0054] The present disclosure performs trajectory prediction by means of sliding inference. The current observation period is a sub-period within the period corresponding to the historical trajectory at the current moment. When the current observation period does not meet the prediction end step condition, the current observation period moves backward step by step according to the preset time step, and the future trajectory of the preset duration is predicted for each observation period. The fused prediction trajectory corresponding to the current observation period moves backward step by step with the current observation period.

[0055] When performing trajectory prediction for the current observation period, not only the historical trajectory of the current observation period is referred to, but also the cross-segment reference information of the previous observation period is combined to ensure the temporal correlation between the front and rear observation periods and improve the coherence and accuracy of the trajectory prediction.

[0056] Among them, the cross-segment reference information is the relevant prediction information corresponding to the previous observation period, which can specifically be the prediction result of the previous observation period, and / or the intermediate state generated during the trajectory prediction process of the previous observation period, such as encoded features, decoded features, etc.

[0057] The utilization method of the cross-segment reference information can be to synchronously input the cross-segment reference information of the previous observation period to jointly carry out trajectory prediction during the process of trajectory prediction; and / or, after obtaining the preliminary prediction result based on the historical trajectory of the current observation period, use the cross-segment reference information of the previous observation period to correct the preliminary prediction result.

[0058] Exemplarily, assume that the length of the historical trajectory at the current moment is T h , and the length of the future trajectory to be predicted is T f , the time length corresponding to the historical trajectory that the model can observe at each step (i.e., the length of the current observation period) T < Th, and the sliding step (i.e., the preset time step) between two adjacent observation periods is t.

[0059] At the first step of inference, the model predicts the trajectory of the future period [T, T + T f seconds according to the historical trajectory of the current observation period [0, T] seconds. The predicted trajectory obtained contains the trajectory that has occurred from T h to T seconds and the trajectory from T + T f to Th The trajectory that did not occur in the second is stored in the memory dictionary, and the cross-segment reference information Y0 generated during the current observation period is stored in the memory dictionary.

[0060] For the i-th step, the model uses the trajectory of the current observation period [t×i, t×i+T] seconds and the cross-segment reference information Y from the previous observation period. i-1 Predict future time periods [t×i+T, t×i+T+T] f The trajectory is measured in seconds, and the predicted trajectory includes T. h The trajectory that has occurred in –t×i seconds and T+T f +t×i–T h The trajectory that did not occur in the second, and the cross-segment reference information Y generated during the current observation period. i Updated to the memory dictionary.

[0061] The prediction termination condition is used to determine whether to terminate the segment-by-segment iterative prediction process. If the prediction termination condition is met in the current observation period, there is no need to proceed to the next observation period. At this point, the fused prediction trajectory corresponding to the current observation period has already incorporated cross-segment reference information from the previous observation period during the segment-by-segment prediction process of each previous observation period. Through multiple iterations, the temporal correlation and information fusion between the preceding and following observation periods are continuously achieved. Determining this fused prediction trajectory as the final trajectory prediction result effectively avoids the error accumulation problem caused by a single prediction, enhances the adaptive response capability to dynamic changes in the agent's trajectory, and improves the accuracy of the trajectory prediction result. Determining the vehicle's driving strategy and controlling its movement based on this trajectory prediction result effectively improves the reliability and safety of vehicle control.

[0062] In addition, the trajectory prediction method in steps S101-S103 can be directly applied to existing trajectory prediction models without making significant modifications to the original architecture, thereby significantly improving long-term prediction accuracy, prediction stability, and the model's generalization ability and robustness.

[0063] In yet another embodiment provided in this disclosure, the current observation period satisfies the prediction final step condition, including at least one of the following:

[0064] The end of the current observation period coincides with the current time.

[0065] The current observation period has reached the preset number of prediction steps;

[0066] The fusion prediction trajectory corresponding to the current observation period does not overlap with the historical trajectory in time.

[0067] For example, if the current time is 4 seconds, the future trajectory for the next 5 seconds is predicted based on the historical trajectory of the previous 4 seconds. That is, the future trajectory for 4-9 seconds is predicted based on the historical trajectory of 0-4 seconds.

[0068] The current observation period is 3 seconds long, with a preset step size of 0.5 seconds. The prediction process is as follows:

[0069] (1) The current observation period is 0-3s. The future trajectory of 3-8s is predicted by the historical trajectory of 0-3s.

[0070] (2) The current observation period is shifted backward according to the preset step size of 0.5s. The current observation period is 0.5-3.5s. Based on the historical trajectory of 0.5-3.5s and the cross-segment reference information of the previous observation period of 0-3s, the future trajectory of 3.5-8.5s is predicted.

[0071] (3) The current observation period is shifted backward according to the preset step size of 0.5s. The current observation period is 1-4s. Based on the historical trajectory of 1-4s and the cross-segment reference information of the previous observation period of 0.5-3.5s, the future trajectory of 4-9s is predicted. The future trajectory of 4-9s is determined as the trajectory prediction result and output.

[0072] The prediction final step condition can be one of the following:

[0073] The end of the current observation period coincides with the current time 4 seconds later;

[0074] The current observation period has reached the preset prediction step count of 3 steps;

[0075] The fusion prediction trajectory (4-9s) corresponding to the current observation period does not overlap with the historical trajectory (0-4s) in time.

[0076] In another embodiment provided in this disclosure, S101 predicts the trajectory of the future time period corresponding to the current observation period based on the historical trajectory of the current observation period and combined with the cross-segment reference information of the previous observation period, to obtain the fused predicted trajectory corresponding to the current observation period, including:

[0077] SA1. Based on the historical trajectory of the current observation period, predict the trajectory of the future period corresponding to the current observation period to obtain the basic predicted trajectory corresponding to the current observation period.

[0078] SA2. Based on the basic prediction trajectory corresponding to the previous observation period, the basic prediction trajectory corresponding to the current observation period is corrected to obtain the fused prediction trajectory corresponding to the current observation period.

[0079] In this embodiment of the disclosure, the cross-segment reference information comes from the basic predicted trajectory corresponding to the previous observation period. The basic predicted trajectory corresponding to the previous observation period contains information such as the movement trend and movement state of the agent in the previous observation period. By correcting the basic predicted trajectory of the current observation period based on the basic predicted trajectory corresponding to the previous observation period, the temporal connection and deviation calibration of the predicted trajectories of the previous and subsequent observation periods can be realized, ensuring the continuity and consistency of trajectory prediction.

[0080] In another embodiment provided in this disclosure, SA2 corrects the basic prediction trajectory corresponding to the current observation period based on the basic prediction trajectory corresponding to the previous observation period to obtain the fused prediction trajectory corresponding to the current observation period, including:

[0081] SA201: Use the decoded features of the basic predicted trajectory corresponding to the current observation period as the query-side features and the decoded features of the basic predicted trajectory corresponding to the previous observation period as the reference-side features. Perform cross-segment attention calculation to obtain feature information for correcting the predicted trajectory.

[0082] SA202. Based on the feature information used to correct the predicted trajectory, the basic predicted trajectory corresponding to the current observation period is corrected to obtain the fused predicted trajectory corresponding to the current observation period.

[0083] In this embodiment of the disclosure, the decoding feature is an intermediate feature output by the trajectory prediction model in the decoding stage of generating the basic predicted trajectory. It is used to characterize the model's prediction information on the agent's future movement trend and driving mode, and can also be called pattern representation.

[0084] For example, for at least one agent, an initial pattern representation is generated based on the historical trajectory of the current observation period, and multiple differentiated pattern representations (i.e., decoded features) are copied. Feature mapping and trajectory regression are performed on each pattern representation through a feedforward network to output displacement sequences corresponding to multiple future times. The displacement sequences of each future time are accumulated sequentially in time order to obtain the corresponding basic predicted trajectory.

[0085] By using the decoded features of the basic predicted trajectory corresponding to the current observation period as query-side features and the decoded features of the basic predicted trajectory corresponding to the previous observation period as reference-side features, cross-segment attention computation can be performed to achieve dynamic association and weighted fusion of the pattern representations of the previous and next observation periods, capture the temporal dependency between the historical prediction pattern and the current prediction pattern, and thus obtain feature information for correcting the predicted trajectory.

[0086] Based on the feature information used to correct the predicted trajectory, the basic predicted trajectory corresponding to the current observation period is corrected to obtain the fused predicted trajectory corresponding to the current observation period. This can effectively suppress trajectory jitter and mode jump caused by a single prediction, and keep the predicted trajectories of the previous and subsequent observation periods consistent in motion trend.

[0087] Specifically, SA201 uses the decoded features of the basic predicted trajectory corresponding to the current observation period as the query-side features and the decoded features of the basic predicted trajectory corresponding to the previous observation period as the reference-side features, performs cross-segment attention calculation, and obtains feature information for correcting the predicted trajectory, including:

[0088] (1) Perform linear mapping on the query-side features and the reference-side features respectively to obtain the corresponding query, key and value, where the query is generated by the query-side feature mapping and the key and value are generated by the reference-side feature mapping;

[0089] (2) Perform a dot product operation on the generated Query and Key to obtain the correlation score between the query-side features and the reference-side features, and then scale the correlation score.

[0090] (3) Mask invalid positions in the correlation score (such as feature positions that are outside the preset observation range or do not have effective motion information);

[0091] (4) Perform softmax normalization on the relevance score after masking to obtain the attention weight, which is used to characterize the degree of influence of different dimensions of information in the reference side features on the query side features;

[0092] (5) The obtained attention weights and Value are weighted and summed to obtain the preliminary fusion representation of the query side features and the reference side features, and to complete the cross-segment information interaction of the decoded features in the observation periods before and after;

[0093] (6) Perform residual connection operation on the preliminary fusion representation to retain the effective information of the original query side features, then perform normalization to eliminate the influence of feature distribution differences, and finally process and optimize the fusion features through a feedforward network to output feature information for correcting the predicted trajectory.

[0094] In another embodiment provided in this disclosure, SA202 corrects the basic predicted trajectory corresponding to the current observation period based on the feature information used to correct the predicted trajectory, to obtain the fused predicted trajectory corresponding to the current observation period, including:

[0095] Through a feedforward network, the feature information used to correct the predicted trajectory is regressed into the correction amount of each position in the basic predicted trajectory corresponding to the current observation period.

[0096] For each position in the basic predicted trajectory corresponding to the current observation period, the correction amount of that position is superimposed on the original coordinates of that position to obtain the fused predicted trajectory corresponding to the current observation period.

[0097] In this embodiment of the disclosure, the feature information used to correct the predicted trajectory is an abstract feature obtained after cross-segment attention fusion. It does not directly correspond to the physical coordinates of the trajectory and cannot be directly applied to the coordinate adjustment of the basic predicted trajectory.

[0098] By using a feedforward network to regress the abstract feature into correction values ​​for each trajectory position, the mapping from abstract features to specific physical adjustment values ​​can be completed. This can be used to correct the temporal correlation and deviation information contained in the feature information of the predicted trajectory, and transform it into a directly calculable correction value, namely the coordinate offset.

[0099] In another embodiment provided in this disclosure, after SA1 predicts the trajectory of future periods corresponding to the current observation period based on the historical trajectory of the current observation period to obtain the basic predicted trajectory corresponding to the current observation period, the method further includes:

[0100] Using the real trajectory in the future as a supervision signal, the model parameters used to generate the base predicted trajectory are optimized to minimize the error between the base predicted trajectory and the real trajectory.

[0101] In this embodiment, supervision is applied to the basic predicted trajectory output for each current observation period. By supervising each time step of the predicted sequence, the evolution pattern of the trajectory can be systematically captured, rather than relying solely on the final prediction result for optimization. Furthermore, incorporating historical trajectories into the supervision signal not only improves the efficiency of training data utilization but also enhances the understanding of historical motion patterns and trajectory context information, thereby improving the dynamic perception and memory capabilities of the model used to generate the basic predicted trajectory in long-term time-series predictions.

[0102] In another embodiment provided in this disclosure, after S101 predicts the trajectory of the future time period corresponding to the current observation period based on the historical trajectory of the current observation period and combined with the cross-segment reference information of the previous observation period to obtain the fused predicted trajectory corresponding to the current observation period, the method further includes:

[0103] By using the real trajectory corresponding to the future time period as a supervision signal, the model parameters used to generate the fused prediction trajectory are optimized to minimize the error between the fused prediction trajectory and the real trajectory.

[0104] In this embodiment, supervision is applied to the fusion prediction trajectory output for each current observation period. By supervising each time step of the prediction sequence, the evolution pattern of the trajectory can be systematically captured, rather than relying solely on the final prediction result for optimization. Furthermore, incorporating historical trajectories into the supervision signal not only improves the efficiency of training data utilization but also enhances the understanding of historical motion patterns and trajectory context information, thereby improving the dynamic perception and memory capabilities of the model used to generate the fusion prediction trajectory in long-term time-series prediction.

[0105] In another embodiment provided in this disclosure, S101 predicts the trajectory of the future time period corresponding to the current observation period based on the historical trajectory of the current observation period and combined with the cross-segment reference information of the previous observation period, to obtain the fused predicted trajectory corresponding to the current observation period, including:

[0106] SB1. Based on the historical trajectory of the current observation period, obtain the basic characteristic information of the current observation period;

[0107] SB2. Based on the fused feature information of the previous observation period, the basic feature information of the current observation period is corrected to obtain the fused feature information of the current observation period;

[0108] SB3. Based on the fusion feature information of the current observation period, predict the trajectory of the future period corresponding to the current observation period to obtain the fusion prediction trajectory corresponding to the current observation period.

[0109] In this embodiment of the disclosure, the cross-segment reference information is derived from the fused feature information of the previous observation period. The fused feature information of the previous observation period can be obtained by extracting features from the agent's historical trajectory, motion state, and scene context within the previous observation period, and is used to reflect the agent's motion pattern, scene information, etc., in the previous observation period.

[0110] By correcting the basic feature information of the current observation period based on the fused feature information of the previous observation period, the temporal correlation and calibration of feature information of the previous and subsequent observation periods can be realized, eliminating deviations and anomalies in the basic feature information of the current observation period, and keeping the feature sequence continuous, stable and consistent.

[0111] Trajectory prediction based on the corrected fused feature information can suppress error propagation and trajectory jumps from the feature source, ensuring the consistency and accuracy of trajectory prediction.

[0112] In some embodiments, when performing trajectory prediction for multiple agents, SB1 obtains basic feature information for the current observation period based on historical trajectories during the current observation period, including:

[0113] First, perform multi-layer self-attention calculation on the historical trajectory of each agent to fuse the trajectory information of the agent at different historical moments and obtain a summary representation of the motion state of a single agent.

[0114] Then, the aggregated representations of each agent, lane element features, and scene surface element features are uniformly constructed into the same feature sequence. Multi-layer global self-attention calculation is performed on this sequence to enable the three types of elements—agent, lane, and surface—to pay attention to each other and transfer features, forming a scene-level interactive representation that includes the global scene context and the interaction relationship between multiple elements.

[0115] Based on this scene-level interaction representation, the basic feature information of each agent during the current observation period is obtained.

[0116] In another embodiment provided in this disclosure, SB2 corrects the basic feature information of the current observation period based on the fusion feature information of the previous observation period to obtain the fusion feature information of the current observation period, including:

[0117] The basic feature information of the current observation period is used as the query-side feature, and the fusion feature information of the previous observation period is used as the reference-side feature. Cross-segment attention calculation is performed to obtain the fusion feature information of the current observation period.

[0118] In this embodiment of the disclosure, by using the basic feature information of the current observation period as the query-side feature and the fusion feature information of the previous observation period as the reference-side feature to perform cross-segment attention calculation, a temporal dependency relationship between the previous and subsequent observation periods can be established. This enables dynamic fusion between the fusion feature information of the previous observation period and the basic feature information of the current observation period. The resulting fusion feature information of the current observation period retains the features within the current observation period and inherits the historical features of the previous observation period. This can eliminate local feature bias, avoid abrupt changes in temporal features, and provide a more robust and coherent feature foundation for subsequent trajectory prediction.

[0119] In some embodiments, when predicting the trajectory for the current observation period, both the basic feature information of the previous observation period and the basic predicted trajectory of the previous observation period can be combined simultaneously. Specifically, S101 predicts the trajectory of the future period corresponding to the current observation period based on the historical trajectory of the current observation period and combined with the cross-segment reference information of the previous observation period, to obtain the fused predicted trajectory corresponding to the current observation period, including:

[0120] SC1. Based on the historical trajectory of the current observation period, obtain the basic feature information of the current observation period;

[0121] SC2. Based on the fusion feature information of the previous observation period, the basic feature information of the current observation period is corrected to obtain the fusion feature information of the current observation period;

[0122] SC3. Based on the fusion feature information of the current observation period, predict the trajectory of the future period corresponding to the current observation period to obtain the basic predicted trajectory corresponding to the current observation period.

[0123] SC4. Based on the basic prediction trajectory corresponding to the previous observation period, the basic prediction trajectory corresponding to the current observation period is corrected to obtain the fused prediction trajectory corresponding to the current observation period.

[0124] For specific implementation methods, please refer to the aforementioned steps SA1-SA2 and SB1-SB3.

[0125] For example, Figure 2 A flowchart of a trajectory prediction method provided in this disclosure embodiment is shown below. Figure 2 As shown, in the first step of trajectory prediction, the historical trajectory information of the current observation period is input into the encoder to obtain the encoded features (i.e., basic feature information) of the current observation period. The decoder outputs the decoded features and the basic predicted trajectory based on the encoded features of the current observation period, and stores the encoded features and decoded features of the first step into the memory dictionary. Since it is the first step of trajectory prediction, there is no cross-segment reference information from the previous observation period. Therefore, the basic feature information of the first step is determined as the fusion feature information, and the basic predicted trajectory of the first step is determined as the fusion feature trajectory. In the second step of trajectory prediction, the historical trajectory information of the current observation period and the memory encoded features (fusion feature information of the first step) stored in the memory dictionary are input into the encoder to obtain the encoded features (i.e., fusion feature information) of the current observation period. The decoder outputs the decoded features and the fusion predicted trajectory based on the encoded features of the current observation period, and updates the encoded features and decoded features of the second step into the memory dictionary. This process continues until the current observation period meets the prediction final step conditions, and the fusion predicted trajectory output in the final step is taken as the trajectory prediction result.

[0126] Table 1 compares the performance of a baseline model and an improved model provided in this disclosure on a multi-agent joint trajectory prediction task. Table 2 compares the performance of a baseline model and an improved model provided in this disclosure on a single-agent trajectory prediction task. The baseline model is a trajectory prediction model that predicts the complete future trajectory in one go. The improved model is constructed by adapting the trajectory prediction method proposed in steps S101-S103 of this disclosure to the baseline model while keeping other parameters and settings of the baseline model unchanged. Tables 1 and 2 compare the performance of the baseline model and the improved model on a large dataset to verify the effectiveness of this disclosure. The training set size is 155,992, and the validation set size is 17,303.

[0127] minFDE: The Euclidean distance between the endpoint of the best predicted trajectory and the endpoint of the true trajectory. The best predicted trajectory is the one with the smallest endpoint error among the K predicted trajectories.

[0128] minADE: The average Euclidean distance between the best predicted trajectory and the true trajectory.

[0129] MR: Based on the endpoint error, the ratio of the number of agents in the predicted trajectory that are not within 2.0 meters of the actual trajectory to the total number of agents.

[0130] AvgMinFDE: The minimum of the average final displacement errors of all agents across K predicted "worlds" is considered the optimal world.

[0131] AvgMinADE: The average displacement error of all agents in the optimal world.

[0132] ActorMR: In the optimal world, the ratio of the number of agents with a final displacement error greater than 2 meters to the total number of agents.

[0133] Table 1

[0134]

[0135] Table 2

[0136]

[0137] As shown in Table 1, in the multi-agent joint trajectory prediction task, after adapting the baseline model to the trajectory prediction method proposed in steps S101-S103 of this disclosure, all indicators were optimized. When the number of inference steps was set to 5, AvgMinFDE improved by 10.6% compared to the baseline model, AvgMinADE improved by 11.5% compared to the baseline model, and ActorMR improved by 13.5% compared to the baseline model.

[0138] As shown in Table 2, in the single-agent trajectory prediction task, after adapting the baseline model to the trajectory prediction method proposed in steps S101-S103 of this disclosure, all indicators were optimized. When the number of inference steps was set to 5, minFDE improved by 10.7% compared to the baseline model, minADE improved by 11.4% compared to the baseline model, and MR improved by 18.7% compared to the baseline model.

[0139] Figure 3 This is a schematic diagram of trajectory prediction results for a benchmark model provided in an embodiment of this disclosure. Figure 4 A schematic diagram of trajectory prediction results for an improved model provided in this embodiment of the disclosure is shown below. Figure 3 and Figure 4As shown, both the baseline model and the improved model predict the trajectories of vehicle A and vehicle B under the same scenario and the same historical trajectory.

[0140] like Figure 3 and Figure 4 As shown, compared to the baseline model, the improved model predicts a trajectory for vehicle A that is closer to the actual trajectory, with a more concentrated distribution and less fluctuation, significantly improving the accuracy and stability of the prediction results.

[0141] For vehicle B, such as Figure 3 As shown, the baseline model's predicted trajectory failed to reflect vehicle B's actual right-turn behavior, only outputting a trajectory close to a straight line, demonstrating its limitations in long-term dynamic inference. In contrast, as... Figure 4 As shown, after applying the trajectory prediction method proposed in steps S101-S103 of this disclosure, the improved model can accurately capture the right-turn intention of vehicle B and generate the corresponding trajectory, while maintaining the smoothness and continuity of the trajectory. Therefore, the trajectory prediction method proposed in steps S101-S103 of this disclosure significantly enhances the prediction system's ability to perceive dynamic changes in the trajectory and its adaptive reasoning ability, thereby improving the accuracy and reliability of long-term trajectory prediction.

[0142] Based on the same technical concept, this disclosure provides a vehicle, including:

[0143] Memory, which stores computer programs;

[0144] A processor is used to run a program, wherein the program executes the vehicle control method provided in any embodiment of this disclosure.

[0145] This disclosure provides a computer device, including a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the vehicle control method provided in any embodiment of this disclosure.

[0146] Based on the same technical concept, this disclosure also provides an electronic device. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure, such as... Figure 5As shown, the electronic device 500 includes a processor 501, a memory 502, and a bus 503. The memory 502 stores execution instructions and includes a main memory 5021 and an external memory 5022. The main memory 5021, also called internal memory, is used to temporarily store computational data in the processor 501 and data exchanged with external memory such as a hard disk 5022. The processor 501 exchanges data with the external memory 5022 through the main memory 5021. When the electronic device 500 is running, the processor 501 and the memory 502 communicate through the bus 503, causing the processor 501 to execute the steps of the vehicle control method provided in any embodiment of this disclosure.

[0147] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the vehicle control method provided in any embodiment of this disclosure. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0148] Through the above description of the embodiments, those skilled in the art can clearly understand that the embodiments of this disclosure can be implemented in hardware or by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this disclosure.

[0149] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes in the drawings are not necessarily essential for implementing this disclosure.

[0150] The sequence numbers of the embodiments disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0151] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A vehicle control method, characterized in that, The method includes: For at least one agent, based on the historical trajectory of the current observation period and combined with the cross-segment reference information of the previous observation period, the trajectory of the future period corresponding to the current observation period is predicted to obtain the fused predicted trajectory corresponding to the current observation period. This includes: obtaining basic feature information of the current observation period based on the historical trajectory of the current observation period; correcting the basic feature information of the current observation period based on the fused feature information of the previous observation period to obtain fused feature information of the current observation period; and predicting the trajectory of the future period corresponding to the current observation period based on the fused feature information of the current observation period to obtain the fused predicted trajectory corresponding to the current observation period. The current observation period is a sub-period within the period corresponding to the historical trajectory at the current moment. If the current observation period does not meet the prediction end-step condition, cross-segment reference information for the current observation period is generated, and the observation period is shifted backward according to a preset step size to obtain the next observation period; the current observation period meets the prediction end-step condition, including at least one of the following: the end of the current observation period coincides with the current time; the current observation period reaches a preset prediction step number; the fused prediction trajectory corresponding to the current observation period does not overlap with the historical trajectory in time; If the current observation period meets the prediction final step condition, the trajectory prediction result of the at least one agent is determined as the fusion prediction trajectory corresponding to the current observation period. Based on the trajectory prediction results of at least one intelligent agent, the vehicle's driving strategy is determined and the vehicle's driving is controlled.

2. The method according to claim 1, characterized in that, The process of predicting the trajectory of future periods corresponding to the current observation period based on historical trajectories of the current observation period and combining cross-segment reference information of the previous observation period to obtain the fused predicted trajectory corresponding to the current observation period includes: Based on the historical trajectory of the current observation period, the trajectory of the future period corresponding to the current observation period is predicted to obtain the basic predicted trajectory corresponding to the current observation period. Based on the basic prediction trajectory corresponding to the previous observation period, the basic prediction trajectory corresponding to the current observation period is corrected to obtain the fused prediction trajectory corresponding to the current observation period.

3. The method according to claim 2, characterized in that, The process of correcting the basic prediction trajectory for the current observation period based on the basic prediction trajectory corresponding to the previous observation period to obtain the fused prediction trajectory for the current observation period includes: The decoded features of the basic predicted trajectory corresponding to the current observation period are used as query-side features, and the decoded features of the basic predicted trajectory corresponding to the previous observation period are used as reference-side features. Cross-segment attention calculation is performed to obtain feature information for correcting the predicted trajectory. Based on the feature information used to correct the predicted trajectory, the basic predicted trajectory corresponding to the current observation period is corrected to obtain the fused predicted trajectory corresponding to the current observation period.

4. The method according to claim 3, characterized in that, The step of correcting the basic predicted trajectory corresponding to the current observation period based on the feature information used to correct the predicted trajectory, to obtain the fused predicted trajectory corresponding to the current observation period, includes: The feature information used to correct the predicted trajectory is regressed into the correction amount of each position in the basic predicted trajectory corresponding to the current observation period through the feedforward network. For each position in the basic predicted trajectory corresponding to the current observation period, the correction amount of that position is superimposed on the original coordinates of that position to obtain the fused predicted trajectory corresponding to the current observation period.

5. The method according to claim 2, characterized in that, After predicting the trajectory of future periods corresponding to the current observation period based on the historical trajectory of the current observation period to obtain the basic predicted trajectory corresponding to the current observation period, the method further includes: Using the actual trajectory of the future time period as a supervision signal, the model parameters used to generate the basic predicted trajectory are optimized to minimize the error between the basic predicted trajectory and the actual trajectory.

6. The method according to claim 1, characterized in that, The process of correcting the basic feature information of the current observation period based on the fused feature information of the previous observation period to obtain the fused feature information of the current observation period includes: Using the basic feature information of the current observation period as the query-side feature and the fusion feature information of the previous observation period as the reference-side feature, cross-segment attention calculation is performed to obtain the fusion feature information of the current observation period.

7. The method according to claim 1, characterized in that, After predicting the trajectory of future periods corresponding to the current observation period based on historical trajectories of the current observation period and combining cross-segment reference information of the previous observation period to obtain the fused predicted trajectory corresponding to the current observation period, the method further includes: Using the real trajectory corresponding to the future time period as a supervision signal, the model parameters used to generate the fused predicted trajectory are optimized to minimize the error between the fused predicted trajectory and the real trajectory.

8. A vehicle, characterized in that, include: Memory, which stores computer programs; A processor for running the program, wherein the program, when running, performs the vehicle control method as described in any one of claims 1-7.

Images