Detection and classification of vehicles for urban traffic scenes

This paper presents a vehicle detection and classification system for urban traffic scenes. This aims to guide surveillance operators and reduce human resources for observing hundreds of cameras in urban traffic surveillance. We perform per frame vehicle detection and classification using 3D models. Motion silhouettes are extracted and compared to a projected model silhouette to identify the ground plane position and class of a vehicle. The system is evaluated with the reference i-LIDS dataset from the UK Home Office. Three weather conditions are tested where sunny conditions give the best classification result of 100% outperforming overcast conditions. The full system including detection and classification for all data achieves a recall of 90.4% at a precision of 87.9% outperforming similar systems in the literature. The i-LIDS dataset is available to other researchers to compare with our results. We conclude with an outlook to use local features for improving the classification and detection performance.