Research On Moving Object Detection Algorithm Based On GMM For Intelligent Video

Intelligent video surveillance technologies, which originate from computer vision, have gradually become the core part of the security systems that are widely used today. Moving object detection, which is the key technology in intelligent video surveillance system, is the foundation of all the high level intelligent analysis and processing such as tracking, behavior understanding and so on. So moving object detection algorithm directly determines the intelligent degree of a surveillance system.Based on the study on the present status of moving object detection and the history of video surveillance technologies, the existing problems of traditional moving object detection algorithm are analyzed in detail. Considering the complexity of monitoring scene, the Gaussian Mixture Model (GMM) was researched, and a new improved algorithm was proposed. The new method improved the completeness and accuracy of motion detection, and enhanced the adaptability of Gaussian model to environmental changes. The main work of this thesis is as follows:(1) The common technologies of moving target detection were studied. The research mainly included image color space conversion, the previous and follow-up process of motion detection and the common motion detection methods. The principle and process of frame difference method and background subtraction method were given. Meanwhile, the comparison and analysis of experimental results verified the characteristics and application scenarios of these two methods.(2) According to the fact that the pixels obey to the Gaussian distribution in the time domain, the motion detection algorithm based on Single Gaussian Model (SGM) and GMM was researched. The specific procedure of background establishment, background updating and motion detection of the two Gaussian models in the entire is given. The experimental results showed that the GMM could get more accurate motion detection in the complex scenes.(3) A new moving object detection method was proposed, which utilized sliding window technology to retain short-term historical information. To a certain extent, the method remedied a fatal shortcoming of the traditional GMM:after a period of modeling, the updating speed of model is difficult to keep up with the real-time changes of true background, which would lead to the increase of false alarm rate in moving object detection. And the new method improved the integrality of motion detection. Furthermore, it reduced the algorithm sensitivity to the scene illumination changes. The experimental results showed that the proposed algorithm could detect moving targets more accurately and perfectly, and better adapted to variations in the environment.(4) In this thesis, it was exploited that the moving shadows couldn’t affect the establishment and update of background model. Firstly, the method used the background subtraction to identify the foreground region. Then combined with the shadow suppression algorithm based on color space, the shadows were separated from foreground region. Finally, the accurate moving objects could be obtained. The method solved the problem that motion detection results often included moving cast shadows because of the light effect to monitoring scene. Comparative experiments with a variety of background subtraction were carried out in the same scene. The results showed that, the algorithm had a better shadow suppression effect when the foreground region was the result of moving target detection based on sliding window GMM.