Lane-Change Detection From Steering Signal Using Spectral Segmentation and Learning-Based Classification

In order to formulate a high-level understanding of driver behavior from massive naturalistic driving data, an effective approach is needed to automatically process or segregate data into low-level maneuvers. Besides traditional computer vision processing, this study addresses the lane-change detection problem by using vehicle dynamic signals (steering angle and vehicle speed) extracted from the CAN-bus, which is collected with 58 drivers around Dallas, TX area. After reviewing the literature, this study proposes a machine learning-based segmentation and classification algorithm, which is stratified into three stages. The first stage is preprocessing and prefiltering, which is intended to reduce noise and remove clear left and right turning events. Second, a spectral time-frequency analysis segmentation approach is employed to generalize all potential time-variant lane-change and lane-keeping candidates. The final stage compares two possible classification methods-1) dynamic time warping feature with k-nearest neighbor classifier and 2) hidden state sequence prediction with a combined hidden Markov model. The overall optimal classification accuracy can be obtained at 80.36% for lane-change-left and 83.22% for lane-change-right. The effectiveness and issues of failures are also discussed. With the availability of future large-scale naturalistic driving data, such as SHRP2, this proposed effective lane-change detection approach can further contribute to characterize both automatic route recognition as well as distracted driving state analysis.