Predicting driver lane change intent using HCRF

Abstract

Accurately predicting drivers intent in advance could help ADAS reduce false alarm rate and improve performance. In this paper, we propose a driver intent prediction approach base on Hidden Conditional Random Fields model. The work can substantially utilize multiple dynamic characteristics of the driving signals, such as the steering wheel angle, lateral position, and drivers gaze compared with other batch process algorithm like Support Vector Machine (SVM). Moreover, it is more discriminative than traditional methods based on Hidden Markov Model (HMM). The experiments were carried out in a driving simulator, and we designed a more complex driving environment compared with previous works. In our experiment, the curvature of the road was not constant and the subjects could make lane change decision on their own. The results show that the proposed method outperforms over SVM and HMM. The prediction accuracy is 99% in 0.5s before the lane change, and 85% in 2s before the maneuver.