Vehicle Detection Based On A Multi-channel Image Feature Pyramid

Vehicle detection technology is an important part of automatic driving system and traffic monitoring system. For the road traffic vehicle detection problem in complex background situations, we researched two aspects-the image feature extraction and classification of detection. Most existing appearance-based vehicle detection systems make use of local features, such as Haar-like, histograms of oriented gradients (HOG), scale-invariant feature transform (SIFT), Harris feature and so on. These local features are still limited to dealing with the illumination, scale, shape variations. A good feature of a vehicle is desired to be discriminative and robust. In this work, in order to achieve the vehicle detection under complex background, we propose a novel vehicle detection based on a multi-channel feature pyramid. Frist, we use scene classification feature to draw out the area which may have vehicle information. Then, in the area which has been selected, we build up a multi-channel feature pyramid by LUV color space, gradient information and context-aware structure features. The context-aware multichannel feature pyramid is able to provide more discriminative and robust feature for vehicle detection. Through the analysis of the current classification algorithms, we finally choose AdaBoost classification algorithm for the vehicle detection system.The results for two public traffic analysis datasets show that the combination of local features and context-aware structure features enhances the identification of vehicle detection system. Moreover, the experimental results are satisfactory even for images containing vehicles which have been undergone scale variations, camera viewpoint change, as well as for images that were photographed under complex background. Our proposed approach leads to better performance than the current state-of-the-art methods. With the research of this paper, we enrich the image detection and recognition technology, and advance the practical process of vehicle detection system.