| As autonomous driving technology gradually landed and was applied,the coexistence of autonomous and human-driven vehicles on the roads was inevitable,which brought severe challenges to vehicle driving safety.In this dynamic traffic environment,identifying the lanechanging intentions of surrounding vehicles and predicting their driving trajectories are the research focus of autonomous vehicles’ perception and understanding of traffic environments and ensuring safe driving.Existing research often overlooks the spatiotemporal interaction between vehicles in dynamic traffic environments or insufficiently extracts vehicle interaction features,leading to limited accuracy in intention recognition and trajectory prediction.This article conducted research on the recognition of lane-changing intentions and trajectory prediction of vehicles in highway scenarios to improve the driving safety of autonomous vehicles.The specific contents are as follows:(1)Analysis and Trajectory Data Processing of Lane Changing Behavior of Highway Vehicles.The characteristics of lane-changing behavior and types of vehicles on highways were studied,and the features of the three stages in the lane-changing process were analyzed.The traffic characteristics between vehicles and surrounding vehicles were qualitatively analyzed,and various factors affecting vehicle lane-changing were investigated.The NGSIM dataset was analyzed in detail,and the existing error problems were intensely studied.The s EMA method and multi-step filtering method were used to clean the abnormal values and noise in the lateral and longitudinal trajectory data of vehicles,reconstructing the velocity and acceleration to provide a more accurate and reliable data foundation for subsequent models.(2)The vehicle lane-changing intention recognition model was constructed.A convolutional residual bidirectional long short-term memory(Bi LSTM)recognition model based on a fusion attention mechanism is proposed to improve the accuracy and prediction capability of vehicle lane-changing intention recognition.The model utilized a one-dimensional convolutional neural network to extract the latent features of the vehicle motion state and constructed the features into a serialized feature vector as the input information of the Bi LSTM network.The optimization bottleneck and gradient vanishing problem of the multi-layer Bi LSTM network were addressed by using residual connections of the deep residual network.The attention mechanism was used to adjust the weights of the residual Bi LSTM network outputs at different time steps,and the Softmax function was applied to calculate the probability of driving intention.The effectiveness of the model was verified using the NGSIM highway dataset.The analysis of overall recognition accuracy and recognition accuracy at different prediction times for lane-changing intention demonstrated that the model had better intention recognition accuracy and prediction capability.(3)The vehicle trajectory prediction model was constructed.To fully utilize the spatiotemporal interaction information between vehicles and improve the prediction accuracy of vehicle trajectories,an encoder-decoder model based on the spatiotemporal attention mechanism was proposed.The model effectively measured the influence of surrounding vehicles on the target vehicle through a spatial attention mechanism,fully describing the spatial interaction relationship between vehicles at different time steps.The temporal attention mechanism was used to capture the dynamic features of vehicle spatial interaction over time.The time features and spatial interaction features at different time steps were fused,and multimodal prediction of vehicle trajectories was achieved by combining features of different driving intentions.The model was evaluated using the NGSIM highway dataset and compared with other models.It was found that the model had the highest prediction accuracy and high computational efficiency,meeting the real-time requirements. |
PDF Full Text Request