Research On Vehicle Trajectory Self-Supervised Learning And Prediction Based On Generative Adversarial

With the continuous development of artificial intelligence algorithms,great progress has been made in autonomous driving technology,and massive trajectory datasets provide data support for autonomous driving technology.The existing high-quality and annotated trajectory datasets are few,and the annotation cost is too high to be applied on a large scale.The correctness of the label information greatly affects the performance of the supervised learning model.Vehicle trajectory prediction can improve the safety of autonomous vehicles.Existing vehicle trajectory prediction models are constructed by data-driven method.As the length of historical trajectory sequences increases,there is a problem of loss or coverage of important feature information,and most vehicle trajectories predictive models neglect to model interactions between vehicles.In response to the above problems,this paper mainly discusses the research on high-quality representation and accurate trajectory prediction of vehicle trajectory data.The main contributions of this paper are as follows:(1)Aiming at the problem that there are few high-quality annotated trajectory datasets and the cost of producing labels is too high,a deep metric temporal neighborhood encoding model is proposed to automatically generate pseudo-labels for trajectory data for representation learning.The trajectory neighborhood representation learning method is used to extract the deep-level trajectory features,and the representation learning of the entire trajectory sequence is obtained by sliding the time window.In addition,according to deep metric learning,the Euclidean distance between benchmark samples,positive samples,and unlabeled samples is constrained to improve the accuracy of the discriminator in identifying neighborhood samples and non-neighborhood samples.Taking the vehicle behavior recognition task as the downstream evaluation task of the model,the model is trained by screening the straight-line and lane-changing data in the High D dataset.A large number of experimental results show that compared with similar models,the model has a better learning effect and an accuracy rate as high as 93.09%,straight-line recall rate of 97.05%,left lane change recall rate of 76.09%,right lane change recall rate of 91.01%,and AUC value of 0.973.(2)Aiming at the problems of missing or covering important trajectory feature information and ignoring the interaction between vehicles,a model based on self-attention social generative adversarial network is proposed to predict vehicle trajectories around autonomous vehicles.Firstly,the self-attention mechanism is used to assign different weights to each feature in the historical vehicle trajectory sequence,so that the model training focuses on these feature information,then the pooling module is used to model the interaction between vehicles,and finally the generative network and the discriminative network are used to train model.The model is trained to deeply fit the distribution of real trajectory data.The US-101 and I-80 datasets in the NGSIM dataset are used to train the model.After a large number of experiments,the prediction accuracy of the model is higher than that of similar algorithms.The errors of ADE and FDE reach 4.97 and 8.92,respectively.