| Unmanned vehicles are the development direction of future intelligent transportation.The development of unmanned driving technology can enhance the traffic capacity of roads and reduce traffic risks.In recent years,unmanned driving technology has developed rapidly,but there is still a lack of research on target vehicle trajectory prediction.Predicting the possible trajectory of the target vehicle in the future can guide the generation of control strategies for unmanned vehicles and greatly improve the driving ability of intelligent vehicles.Therefore,this paper studies the trajectory prediction of vehicles,and the main research contents include the following points:First of all,this paper summarizes the research status of vehicle trajectory prediction at home and abroad,and clearly solves the data problem required for trajectory prediction.This paper proposes to detect the target vehicle by perceiving the surrounding environment,which has practical application value.The RGB camera on the vehicle side is used as the input device,YOLOv5 is used as the target detector,and Deepsort is used as the tracker to track the target vehicle,so as to obtain the historical trajectory information of the vehicle for a period of time.The validity of the algorithm is verified by selecting the driving video in the real scene.After obtaining the trajectory information of the target vehicle,the vehicle trajectory prediction model is designed.In order to overcome the shortcomings of the sequential calculation of the traditional sequence prediction model,this paper proposes a trajectory prediction model TFL-MLP-CSP(Transformer-LSTM-Multilayer Perceptron-Convolutional Social Pooling)by using Transformer as the coding structure.In order to overcome the drawbacks of the original Transformer model that requires multiple decoding,a multi-layer perceptron decoding structure is designed to quickly predict the vehicle trajectory at all times.The convolution social pooling module is added to extract the neighbor vehicle information of the target vehicle to improve the prediction accuracy.The open source dataset NGSIM in the United States is tested,and the dataset is divided into different working conditions,and the prediction effect of the model is compared on the subdivided dataset.In order to verify the effectiveness of the above modules,related ablation experiments and visualization of vehicle trajectory prediction results were studied.Experiments show that TFL-MLP-CSP is 33.2 % ahead of Transformer method in ADE index,35.7 % ahead of FDE index and 33.1 % ahead of RMSE index.Finally,the text data of visual perception contains less information,and the model built based on it has limitations.Therefore,this paper uses binocular cameras and lidar as information sources.In order to overcome the feature drift phenomenon in the encoding and decoding structure,a pyramid feature(FPN)branch is added to the base network.The FPN branch can introduce shallow spatial information into the deep semantic feature layer of the model,which improves the prediction effect of the model.Relevant experiments are carried out on the KITTI dataset to prove the predictive ability of the model.At the same time,the time-consuming analysis of the model was carried out,showing that the FPSP(Feature Pyramid networks based on Skip connect for vehicle Prediction)proposed in this paper has a small amount of calculation compared with the INFER model,and the prediction error on the ADE index is reduced by 31.8%.The prediction on the FDE index is reduced by 39.7%. |
