Research On Foreground Extraction And Targets Detecting And Tracking For Video Surveillance

Nowadays, the social contradictions lead to the frequent occurrence of emergency and terrorist attack with the fast development of society. The city video surveillance system plays an important role in protecting the social safety as one of the fundamental partition of Safe City System. It’s a big challenge to monitor the whole city with manual work, because the video information is huge. So we need an intelligent surveillance system to assist us to obtain and take advantage of the information accurately and swiftly, which can improve the quality of security service. With the improvement of technology and the breakthrough of computer vision, we have already been able to build a large-scale networked digital surveillance system, and it is possible to build a new generation intelligent surveillance system using artificial intelligence.It is a challenging work to research and develop an intelligent surveillance system, which contains many fields knowledge about computer science. Extraction of foreground and the objects tracking, recognition are two core issues in intelligent video surveillance. In this thesis, we study the problems about extracting foreground and object tracking, detecting, and propose some solutions for these problems. The main content of this paper is shown as follow:1. We implement a dynamic background model based on multi-channel Gaussian mix model to extract the foreground of video. To increase the accuracy and decrease the time cost, we propose an improved model for real-time tracking. However, the foreground always contains some noise points and holes inside objects. So we introduce an image morphology method to solve this problem.2. We summarize three general feature models for object recognition, HOG, Haar-like and LBP. Then we do some experiments to detect human face using cascade Adaboost, and detect pedestrians using S VM classifier.3. We also introduce a novel multiple objects tracking strategy. Our tracking framework is based on Lucas-Kanade optical flow tracker. To improve the performance, we use the optical flow of mean sampling points as descriptor and implement a forward- backward error detector. In order to test the performance of the system, we select three real surveillance video for evaluation and obtain good result.4. In order to decrease the time consuming, we take advantage of GPGPU technology(CUDA) to promote the time performance.