Abnormal Moving Vehicle Detection for Driver Assistance System in Nighttime Driving

This paper proposes a new approach of abnormal vehicle detection for frontal and lateral collision warnings in nighttime driving using monocular vision. Motion information is used to estimate moving objects. An empirical threshold range is introduced to eliminate efficiently most of non-vehicle regions. Vehicle candidates are segmented by using K-means clustering. An analysis is performed carefully to consider what initial K value is optimal for vehicle region segmentation. The vehicle candidates are classified by using Support Vector Machine (SVM) classification. The aforementioned method has high ability to retain the abnormal moving vehicles. The detected abnormal vehicles consist of on-coming, overtaking, change speed, change lane, and road-side parking. These vehicles are dangerous with respect to the host vehicle. Experimental results show that the proposal approach is useful for real-time collision warning function of driver assistance system in nighttime driving.