A Deep Learning Framework to Explore Influences of Data Noises on Lane-Changing Intention Prediction

The accuracy of the data is crucial to the real-time prediction of autonomous driving. Due to factors such as weather and the accuracy of data collection equipment, there frequently exist noises in the data collected in real time. Therefore, it is necessary to perform analysis on acquired kinematic features related to driving behavior prediction. This study proposes a novel deep learning framework to explore influences of data noises on lane-changing intention prediction. Kinematic features including the longitudinal distance difference, velocity and acceleration, lateral velocity and acceleration of the vehicles are first extracted from the HighD. Then, the anti-interference performance of deep learning models such as transformer is tested. By comparing dataset with and without noises, we develop an evaluation method containing several predictive performance metrics and statistical measures. The results show that: (1) the longitudinal acceleration of the vehicle has the lowest sensitivity to noise, and the lateral velocity has the weakest anti-interference and the highest sensitivity. (2) The Bi-LSTM model with multi-head attention mechanism performs well in reducing the sensitivity of longitudinal acceleration and prediction accuracy. This study provides valuable information for data acquisition and model selection of real-time driving intention prediction.