Lane Changing Maneuver Prediction by Using Driver’s Spatio-Temporal Gaze Attention Inputs for Naturalistic Driving
Jingyuan Li,Titong Jiang,He Liu,Yingbo Sun,Chen Lv,Qingkun Li,Guodong Yin,Yahui Liu
DOI: https://doi.org/10.1016/j.aei.2024.102529
IF: 8.8
2024-01-01
Advanced Engineering Informatics
Abstract:Driver's lane changing maneuver prediction holds significant importance in enhancing the functionality of the Advanced Driver Assistance System (ADAS) and ensuring driving safety. This study introduces an innovative approach to model the driver's spatio-temporal gaze attention to the traffic environment, and realizes highprecision lane changing prediction by integrating multi-source inputs. Firstly, a naturalistic driving dataset containing information of the ego vehicle, the traffic environment, and the driver's eye movements is created based on the real-vehicle apparatus. Secondly, the driver's gaze attention to the traffic environment is creatively modeled by introducing a gaze cone and quantified by using spatio-temporal indexes. Concurrently, vehicle dynamic inputs and traffic environmental inputs including the information of surrounding vehicles and the lane are also constructed. Thirdly, a Transformer Encoder model, containing the multi-head attention mechanism, is established to extract the information from input sequences for maneuver prediction. Performance comparison is carried out using various historical horizons, input combinations, and prediction horizons. The model shows an enhanced prediction performance after adding gaze attention inputs, with the accuracy, precision, recall, F1score, and Area Under Curve (AUC) being 0.945, 0.937, 0.931, 0.934, and 0.986, respectively. In addition, as the prediction horizon extends from 1 s to 3 s, the model utilizing multi-source inputs consistently demonstrates the highest accuracy and minimal performance decline, exhibiting stable predictive capabilities. In future application, the proposed driver's gaze attention model can serve to alert drivers about unobserved objects, and the maneuver prediction method shows promise for the integration into lane change assistance systems.