A GMM-SVM Approach to Vehicle Type and Color Classification

We describe our approach to segmenting moving road vehicles from the color video data supplied by a stationary roadside CCTV camera and classifying those vehicles in terms of type (car, van and HGV Heavy Goods Vehicle) and dominant color. For the segmentation, we use a recursively updated Gaussian mixture model approach, with a multi-dimensional smoothing transform. We show that this transform improves the segmentation performance, particularly in adverse imaging conditions, such as when there is camera vibration. We then present a comprehensive comparative evaluation of shadow detection approaches, which is an essential component of background subtraction in outdoor scenes. For vehicle classification, a practical and systematic approach using a kernelized support vector machine is developed. The good recognition rates achieved in our experiments indicate that our approach is well suited for pragmatic vehicle classification applications.