Intrusion Detection Algorithm And Dsp Implementation Of Intelligent Video Surveillance

With the increased demand for social public security, video surveillance is constantly extended to every corner of society. The traditional way of video surveillance has been unable to meet the increasing surveillance need. The demand of intelligent video surveillance with all day long, unattended, automatic alarm, real-time intelligence is becoming increasingly urgent.This paper is committed to the research of foreground object detection, video stabilization, object tracking in video surveillance, and have a research on code optimization methods on DM642 DSP device. We implemented these algorithms on the DSP, and construct a robust and real-time intrusion detection system of intelligent video surveillance, which is DSP core based hardware platform.This paper mainly covers the following topics:(1) A common approach for discriminating moving objects from the background is detection by background subtraction. We make a full analysis of the adaptive mixture of Gaussian model and codebook model. Considering the robustness and real-time requirements, we choose codebook model to model the background.(2) We propose a robust real-time video stabilization method. In the proposed algorithm, we first apply fast features detection algorithm to detect features of frames, and compute the sparse optical flow using a hierarchical phase only correlation (POC) tracking algorithm, which can run very fast in DSP based devices. Then, Global motion of frame is estimated using RANSAC algorithm. Finally, the camera motion parameters are smoothed temporally by Gaussian filter method to obtain a stable and smooth camera motion path, and we transform all frames to obtain a stabilized video. Experiment results demonstrate that the presented approach is real-time and robust.(3) In order to track the moving object, we make a full analysis of the mean-shift and particle filter object tracking method, and compare their advantage and drawback. We propose an improved color histogram to describe the object, and make a combination of the mean-shift and particle filter tracking method, which is very fast and robustness. (4) We have a research on code optimization methods on DM642 DSP device, and use those methods to optimize the intrusion detection code we have implemented. We also have a discussion on the design of the system of software framework.