CN-121996942-A - Track prediction method, system, medium and product based on mixed expert model

CN121996942ACN 121996942 ACN121996942 ACN 121996942ACN-121996942-A

Abstract

A track prediction method, system, medium and product based on a mixed expert model relate to the field of data processing systems specially suitable for prediction purposes, and the method comprises the steps of obtaining an observation track with any length; the method comprises the steps of utilizing a preset encoder to encode an observation track to obtain track characteristic representation, inputting the track characteristic representation into a gating network to calculate an optimal observation length, activating a corresponding target expert module in a preset mixed expert model according to the optimal observation length to optimize the observation track to obtain an optimized track, wherein the optimizing process comprises the steps of comparing the current length of the observation track with the optimal observation length, executing observation pruning if the current length is greater than or equal to the optimal observation length, executing reverse prediction if the current length is less than the optimal observation length, and inputting the optimized track into a track prediction backbone network to obtain a future track prediction result. By implementing the method and the device, the track prediction accuracy under the observation condition of any length can be improved.

Inventors

LI CHANGSHENG
SUN JINYUE
LI TAO
LV RUILIN
LI YUHANG
LI BOYANG
YUAN YE
WANG GUOREN
YANG JINGHUI
HU ZENG
ZHANG LIZHI

Assignees

北京理工大学
航天万源云数据河北有限公司
北京理工大学唐山研究院

Dates

Publication Date: 20260508
Application Date: 20251219

Claims (10)

1. A trajectory prediction method based on a hybrid expert model, applied to a trajectory prediction system, the method comprising: Obtaining an observation track with any length; Encoding the observation track by using a preset encoder to obtain a track characteristic representation; Inputting the track characteristic representation into a gating network, and calculating to obtain the optimal observation length matched with the observation track; activating a corresponding target expert module in a preset mixed expert model according to the optimal observation length; The target expert module is used for optimizing the observation track to obtain an optimized track, wherein the optimization comprises the steps of comparing the current length of the observation track with the optimal observation length, executing observation pruning to remove interference information if the current length is greater than or equal to the optimal observation length, and executing reverse prediction to supplement missing historical tracks if the current length is less than the optimal observation length; and inputting the optimized track into a track prediction backbone network to obtain a future track prediction result.
2. The method according to claim 1, wherein the step of encoding the observation trajectory with a preset encoder to obtain a trajectory feature representation comprises: inputting the coordinate sequence of the observation track into a preset encoder, and mapping to obtain a hidden layer sequence; Calculating relative time intervals among all time steps in the observation track, and mapping the relative time intervals into rotary position coding parameters; performing rotation transformation on the hidden layer sequence based on the rotation position coding parameters to obtain an intermediate characteristic sequence carrying relative time sequence information; and executing pooling operation on the intermediate feature sequence to obtain the track feature representation.
3. The method according to claim 1, wherein if the current length is smaller than the optimal observed length, the step of performing backward prediction to supplement the missing historical track comprises: Calculating the absolute value of the length difference between the optimal observation length and the current length to obtain the number of time steps to be supplemented; inputting the track characteristic representation into a reverse prediction network, and generating historical coordinate points with the number equal to the number of the time steps; and splicing the historical coordinate points to the initial end of the observation track in time sequence to obtain a completed track sequence, and outputting the completed track sequence as the optimized track.
4. The method of claim 1, wherein prior to the step of acquiring an arbitrary length of observation trajectory, the method further comprises: Acquiring a complete training sample track in a training set, and performing truncation processing on the complete training sample track to obtain a truncation observation track and a corresponding truncated real historical track; inputting the truncated observation track into a reverse prediction network to be trained, and generating a prediction history track; Calculating a reconstruction loss value between the predicted historical track and the truncated real historical track; The predicted historical track and the truncated observation track are spliced and then input into the track prediction backbone network, so that a predicted future track is obtained; calculating a predicted loss value of the predicted future track and the real future track; And performing gradient update on network parameters of the backward prediction network based on the reconstruction loss value and the prediction loss value.
5. The method of claim 4, wherein after the step of performing a gradient update to network parameters of the reverse predictive network based on the reconstructed loss value and the predicted loss value, the method further comprises: Recording the activation times of each expert module in the mixed expert model in a training batch to obtain an activation times list; Calculating the mean value and standard deviation of each element in the activation times list to obtain a distribution parameter; Calculating to obtain a load imbalance coefficient based on the ratio of the standard deviation to the mean value; And when the load unbalance coefficient exceeds a preset balance threshold, a punishment gradient is applied to the weight parameter of the gating network, and adjustment is performed on the weight parameter of the gating network based on the punishment gradient so as to balance the activation probability of each expert module.
6. The method according to claim 1, wherein after the step of activating a corresponding target expert module in a preset hybrid expert model according to the optimal observation length, the method further comprises: Calculating a length difference value between the current length and the optimal observation length to obtain a length deviation value; when the observation track is determined to be positioned in the classification boundary region based on the length deviation, extracting the activation probability distribution of each expert module output by the gating network, and identifying candidate expert modules with activation probability larger than a preset probability threshold to obtain a candidate expert set; Driving each candidate expert module in the candidate expert set to execute processing on the observation track to obtain a plurality of candidate optimized tracks; According to the activation probability corresponding to each candidate expert module as a weight coefficient, performing weighted fusion on the time-step coordinates of the plurality of candidate optimization tracks after time alignment to obtain a fusion optimization track; and substituting the fusion optimization track for the optimization track, and inputting the track prediction backbone network.
7. The method according to claim 6, wherein after the step of performing weighted fusion on the corresponding time-step coordinates of the plurality of candidate optimized trajectories according to the activation probabilities corresponding to the respective candidate expert modules as weight coefficients to obtain a fused optimized trajectory, the method further comprises: calculating displacement vectors between adjacent time steps in the fusion optimization track to obtain a displacement sequence; Performing second-order difference on the displacement sequence to obtain an acceleration characteristic sequence representing the smoothness of the track; detecting abnormal time steps of the model in the acceleration characteristic sequence exceeding a preset physical limit threshold value, and generating a motion mutation set; And performing smooth correction on coordinate points corresponding to the motion mutation set in the fusion optimization track by using a local polynomial interpolation algorithm to obtain a smooth fusion track conforming to kinematic constraint, and inputting the smooth fusion track as the optimization track into the track prediction backbone network.
8. A trajectory prediction system comprising one or more processors and memory coupled with the one or more processors, the memory for storing computer program code comprising computer instructions that the one or more processors invoke to cause the trajectory prediction system to perform the method of any of claims 1-7.
9. A computer readable storage medium comprising instructions which, when run on a trajectory prediction system, cause the trajectory prediction system to perform the method of any one of claims 1-7.
10. A computer program product, characterized in that the computer program product, when run on a trajectory prediction system, causes the trajectory prediction system to perform the method according to any one of claims 1-7.

Description

Track prediction method, system, medium and product based on mixed expert model Technical Field The application relates to the field of data processing systems specially suitable for prediction purposes, in particular to a track prediction method, a track prediction system, a track prediction medium and a track prediction product based on a hybrid expert model. Background The track prediction technology is used for predicting a future motion path of the traffic participant according to the historical motion track of the traffic participant, and is the basis of the automatic driving system for carrying out path planning and obstacle avoidance decision. The track prediction method based on deep learning has become a mainstream technology in the field by learning a motion pattern through a structure such as a cyclic neural network, a convolutional neural network or a transducer. In the related art, a trajectory prediction method based on deep learning is generally trained using public data sets, in which trajectories have a fixed observation length. The model receives a historical track sequence with fixed length during training, space-time characteristics are extracted by an encoder, and future track prediction is generated by a decoder. Part of the method is introduced into an attention mechanism to capture time dependency relationship or uses a graph neural network to model interaction of a plurality of traffic participants, and the training mode based on fixed-length input can obtain better prediction accuracy on a standard data set. However, the length of the historical track acquired by the system varies due to factors such as occlusion, limited field of view of the sensor, abrupt appearance of the target, or interruption of tracking. Too short observation sequences cause the model to lack enough time sequence information and have larger prediction deviation, and too long observation sequences exceed the length distribution during training, so that a position coding and time sequence modeling mechanism is disabled, and atypical motion modes in the model can interfere model judgment, and the prediction accuracy is reduced. Disclosure of Invention The application provides a track prediction method, a system, a medium and a product based on a hybrid expert model, which are used for improving the track prediction accuracy under the observation condition of any length. The track prediction method based on the mixed expert model is applied to a track prediction system, and comprises the steps of obtaining an observation track with any length, encoding the observation track by a preset encoder to obtain track characteristic representation, inputting the track characteristic representation into a gating network to calculate to obtain an optimal observation length matched with the observation track, activating a corresponding target expert module in the preset mixed expert model according to the optimal observation length, optimizing the observation track based on the target expert module to obtain an optimized track, wherein the optimizing process comprises the steps of comparing the current length of the observation track with the optimal observation length, executing observation pruning to remove interference information if the current length is greater than or equal to the optimal observation length, executing reverse prediction to supplement the missing historical track if the current length is smaller than the optimal observation length, and inputting the optimized track into a track prediction backbone network to obtain a future track prediction result. In the above embodiment, the track prediction system dynamically calculates the optimal observation length for each input track with any length by using the encoder and the gate control network by adopting the above technical scheme, activates a specific expert module to perform differentiation processing based on the comparison result of the optimal length and the current track length, so that the quality of data input to the final prediction network is improved, the problem of performance degradation when track observation inconsistent with training data distribution is processed in the related technology is solved, and track prediction accuracy is enhanced. With reference to some embodiments of the first aspect, in some embodiments, the step of encoding the observation track by using a preset encoder to obtain a track feature representation specifically includes inputting a coordinate sequence of the observation track into the preset encoder, mapping to obtain a hidden layer sequence, calculating a relative time interval between time steps in the observation track, mapping the relative time interval into a rotation position encoding parameter, performing rotation transformation on the hidden layer sequence based on the rotation position encoding parameter to obtain an intermediate feature sequence carrying relative time sequence information, and performi