A Comparative Study of Different Features for Vehicle Classification

This paper presents a comparative study of different features for vehicle classification. Real-time vehicle classification system using computer vision is relatively cheaper and easy to install. As traffic is heterogeneous in India, road planning and traffic management is challenging. So an automated vehicle detection and classification system is useful for traffic survey, planning, signal time optimization and surveillance. In this work, traffic video data is collected using a camera placed on the top of a vehicle parking on the side of a road at an angle of approximately 45°. Both audio and video are used for vehicle detection. The presence of a vehicle is detected from frames corresponding to the peaks in the short time energy of audio. The process of adaptive background subtraction is performed on the selected frames to separate the vehicle from the background. After background subtraction, morphological processes such as erosion, dilation and closing are applied to get the region of interest. There may be mulitiple frames with the same vehicle are detected at this stage. To reduce the multiple occurrences of the same vehicle in selected frames, Speeded-Up Robust Feature (SURF) matching algorithm is used. Different features like Histogram Oriented Gradient (HOG), Local Binary Pattern (LBP), KAZE, Binary Robust Invariant Scale Keypoint (BRISK) features of selected frames are extracted and Support Vector Machine (SVM) models are developed. Vehicle classification accuracy of various features are compared using a 20 minutes traffic video. It is observed that HOG gives the best result compared to KAZE, LBP and BRISK, with an accuracy of 85.50%.