Research On Key Technologies Of LSTM-based Driver Intention Recognition In High-speed Environments

With the continuous improvement and expansion of highways in China,the issue of traffic safety has become increasingly prominent,with a sharp increase in the incidence of accidents,attracting widespread attention.In recent years,the emergence of autonomous vehicles has provided a new opportunity to effectively reduce traffic accidents caused by human driving errors,offering a solution to the problem of highway traffic safety.Driver intention recognition is a key technology in autonomous driving systems,as identifying driver intentions in advance can assist drivers in making driving decisions and significantly reduce traffic accidents caused by lane changes.This thesis addresses the pain points of frequent accidents on highways and the need for improved driver intention recognition technology by using a modified LSTM network to conduct systematic research in two areas: driver intention recognition and vehicle trajectory prediction based on driver intentions,validated through experiments on the NGSIM dataset.Specifically,the thesis achieved the following results:(1)Addressing the problem of interference from outliers in the dataset affecting experimental results,the thesis identified and corrected outliers using boxplots and applied a symmetric moving average algorithm to smooth vehicle trajectory data.The results showed that data cleaning improved the quality of vehicle trajectory data while retaining the original fluctuation characteristics of the data.The thesis also used a method based on specific rules to select five types of driving intention data from the dataset,laying the foundation for data training and validation in subsequent sections.(2)Addressing the problem of low driver intention recognition accuracy and insufficient utilization of surrounding contextual information,the thesis constructed a CNN-LSTM-Attention driver intention recognition model that optimized the optimization of neighbor information of surrounding vehicles and self-features in the feature layer,a factor overlooked in most studies.The thesis quantitatively analyzed the appropriate historical sequence length required by the model,and validated the impact of prediction time on driver intention recognition indicators through analysis of different arrival times at lane-changing points.Comparative experiments and ablation experiments on the NGSIM dataset were conducted to explore the performance of the network model in driver intention recognition,with the results showing that the proposed driver intention recognition model outperformed other prediction models in all indicators.(3)Addressing the problem of a lack of analysis of driver behavior in vehicle trajectory prediction,the thesis constructed a Conv LSTM-MDN encoding-decoding vehicle trajectory prediction model based on driver intention recognition results.Using RMSE as the evaluation index,the thesis clearly demonstrated that driver intention recognition results significantly improved the accuracy of vehicle trajectory prediction.Compared to the Conv LSTM model in trajectory prediction experiments,the experimental results showed that the proposed model could simultaneously output multiple different trajectory prediction modes,with better multi-modal prediction capability.Ablation experiment results showed that the Conv LSTM-MDN model incorporating driver intention recognition had a good fitting effect on time-series trajectory prediction and could more effectively achieve the trajectory prediction of surrounding traffic vehicles.