Abstract:

In an autonomous driving system, the system needs to accurately identify the driver's intention to help them drive safely in complex traffic scenarios. Aiming at the issue of low accuracy of driver intent recognition and lacking consideration of the influence of optimized features on the accuracy of the model currently, a driver intention recognition method based on a time series model is proposed using deep learning knowledge. The method is based on Attention mechanism and incorporates Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM), introducing own and environmental information of the vehicle as spatiotemporal input to capture the spatial interaction and temporal evolution of surrounding vehicles. The method can simultaneously predict the driver intention in both lateral and longitudinal directions of the target vehicle and is trained and verified on the actual road dataset next generation simulation. The experimental results show that the proposed CNN-LSTM-Attention model can accurately predict the driver's driving intention in the highway environment, which has obvious advantages over the LSTM and CNN-LSTM model and provides an effective guarantee for the safe operation of the automatic driving system.

Key words: automatic driving, intention identification, Attention mechanism, convolutional neural network, long short-term memory network