Video-based Vehicle Detection and Classification in Heterogeneous Traffic Conditions using a Novel Kernel Classifier

Vehicle classification in a traffic video is considered a difficult task due to similarity in appearances among different vehicles. This paper presents a real-time algorithm for detection and classification of different categories of vehicles in a heterogeneous traffic video. The processing of the video is done in four steps starting with camera calibration, vehicle detection, speed estimation, and classification. Vehicle detection is achieved by using background subtraction and blob tracking method. Speed of the detected vehicle is estimated by utilizing virtual start and stop line markers and calibration parameters. The vehicle classification is done by extracting multiple features of the detected vehicles which serve as input to a support vector machine based classifier. A histogram-based nonlinear kernel is used in the classifier. A combination of interest point detectors and low-level shape detectors as features was found to produce accurate and consistent results.