Vehicle classification approach based on the combined texture and shape features with a compressive DL

Abstract

Automatic vehicle classification is a fundamental task in intelligent transportation systems. Image-based vehicle classification is challenging due to occlusion, low-illumination, and scale change. This study proposes an innovative approach by combining texture and shape features into a complementary feature and using the complementary feature to train a compressive dictionary to improve accuracy and efficiency. In the feature combination, the scale-invariant feature transform descriptor is applied to extract the texture and shape features from the original vehicle images and their edge images, respectively. In the dictionary training, a compressive dictionary learning (DL) algorithm, called compressive K singular value decomposition (CKSVD) algorithm, is proposed to improve the dictionary training efficiency. The CKSVD algorithm divides the feature dictionary into several same-sized data blocks and then performs the DL in each data block based on a very sparse random projection matrix. On the feature combination and DL, the proposed approach employs the kernel sparse representation method to classify vehicles into four types: buses, trucks, vans, and sedans. The kernel sparse representation method enables the linearly inseparable classification in the combined feature space to be linearly separable. Experimental results show that the proposed approach can improve the accuracy and efficiency of vehicle classification.