A Research On Vehicle Detection Technology Oriented To Surveillance Video Application

Object detection and recognition is the key to intelligent transportation and unmanned technologies. Due to immature technology, Object detection and recognition in the current traffic video surveillance system has not been widely applied, the relevant theory and technology has been of great challenge, but also has greatly research significance.The main task of video-based target detection and recognition is to find an object of interest in video frames, involving moving object detection, tracking and classification. In this paper, several video object detection and classification tools based on traditional methods such as HOG is implemented, and sparse coding and its application in target recognition is studied.Firstly, we study the recognition techniques of vehicle detection and recognition based on traditional features. Detection accuracy and performance of frame difference and background modeling is compared. We also implemented a tool to detect a given target in a video by using histogram of colors and SIFT feature. Classification based on HOG feature and SVM classifier is studied as well.Secondly, we study dictionary training techniques in sparse coding. Kinds of dictionary training methods is studied and implemented, including L2-sparse-coding, L1-soft-thresholding, and L1-approximation. We propose a new training method, which combining L2-sparse-coding and L1-approximation. Experiments show that the proposed method improves performance compared with L1-approximation, and achives much better dictionary than L2-sparse-coding.Finally, we studied the hierarchical sparse coding techniques. We designed and implemented shallow and middle level hierarchical sparse coding network. And vehicle classification experiments conducted in the shallow sparse coding network show better accuracy than HOG. Experiment shows that overlapped pooling method outperforms the non-overlapped one. And a theoretical analysis of the invariance of the middle level hierarchical sparse coding is also presented in this paper.