Moving Object Detection Under Dynamic Background Based On Codebooks

Video surveillance has been widely applied in the modern world, thanks to the powerful computing and storage techniques. Especially in this world of 3G telecommunication, users can access their surveillance cameras remotely at any time. While existing techniques mainly focus on the static background, they are weak when facing the dynamic background. Therefore we are interested in this dynamic background problem. Particularly in this thesis we study the moving object detection under the dynamic background using codebooks.The main work of the research includes:(1) Studying the basic approaches existing to detection moving objects:We summarize existing methods, analyze their advantages and disadvantages. We mainly introduce the pipeline of background-subtraction model and corresponding methods.(2) Presenting the pixel-based HSV codebook model:HSV color space which has more distinguish power than RGB color space is introduced., This model is fast with low computational complexity by Reducing the size of codewords in the former method from 9 to 4 and incorporating with a simply method to compute color and illumination distortions. The model also adopts a non-reduction codebook for background modeling because of the high dynamic and long-lasting properties of the background, which ensures few holes detected.(3) Introducing the region-based HSV codebook model:Our research shows that region features are generally more stable than pixel-based features. In this region-based model, a threshold factor is proposed so that the threshold will not change significantly when the background changes drastically. The model also uses small training set and thus accelerates the detection process. Furthermore, a feedback-based correction method is included to remove false targets during learning. At last, two new evaluation curves, recall curve and precision curve, are introduced to measure the performance globally.