Driving Intention Recognition Model Based on Bi-GLSTM Network

Driving intention recognition is promised to effectively improve the vehicle’s ability to predict the trajectory of other traffic participants. However, the interaction of surrounding vehicles in a dynamic and complex traffic environment is one of the most challenges to be solved. In order to improve the accuracy of driving intention recognition in dynamic and complex traffics, a time-series recognition model of driving intention based on Bi-GLSTM network is proposed. First, the position, velocity and acceleration in the original dataset are smoothed based on the Local weighted regression scatter smoothing method. Also, the driving datasets are labeled with driving intention based on the longitudinal and lateral motion parameters. Subsequently, a graph attention neural network is established to extract interaction features among surrounding vehicles, where attention mechanism is embedded, to enhance highly related vehicle motion states. To further improve the robustness of the model in dynamic and complex traffics, Bidirectional long short-term memory network is used to extract deep temporal features of interaction and historical motion. Moreover, our model is trained and verified on the public datasets HighD. Compared with GNN and RNN model, the recognition accuracy increased by 11.33%, 55.31%. By visualizing the attention weight, it shows that the proposed model also solves the problem of explainability to a certain extent. © 2024 Chinese Mechanical Engineering Society. All rights reserved.