Research On Moving Object Detection Algorithms In Intelligent Visual Surveillance System

According to the development of surveillance technology, surveillance technology can be approximately divided into three stages: human site surveillance, human video surveillance and intelligent visual surveillance. Currently, governments and scholars around the world are paying their close attention to a new generation of surveillance technology—intelligent visual surveillance technology. Intelligent visual surveillance technology is a multi-disciplinary compositive issue, and is also a challenging frontier topic, which is involved in image processing, image analysis, machine vision, pattern recognition, artificial intelligence and many other research fields. It is widely used in various application fields and has great potential in market value and social benefit. The process of intelligent analysis of visual surveillance consists of four basic steps: moving object detection, classification, tracking and video analysis, where moving object detection is the foundation. Due to dynamic changes of the background image, such as weather, light, shadow and background interference, moving object detection becomes a quite difficult task. Therefore, moving object detection in intelligent visual surveillance system is of great significance.This topic comes from Science and Technology Key Project of Shandong Province--"Intelligent Visual Surveillance System", researches, designs and develops an intelligent visual surveillance system. We program the intelligent visual surveillance system by using Visual C++, HaiKang SDK, SQL Server 2000 etc based on HaiKang DS-4004H board. The system is composed of six individual parts: central database, local database, data management center, video center, surveillance center and streaming media client. It achieves a simple multi-media intrusion detection incorporating sensor and camera in video server by using multi-media surveillance system for reference.This paper introduces some basic methods of moving object detection: optical flow method, temporal difference method and background subtraction method, analyses and compares each method, and indicates their advantages, disadvantages and application ranges. Improved methods of moving object detection are classified into three categories. A fast improved algorithm of Gaussian mixture model is presented for application of visual surveillance with a static camera. By function analysis of Gaussian variance and mean value, and considering the complexity of the learning rate of mean value, the updating equation of variance is omitted. Via experiments under different variances, we confirm the appropriate value of variance, and a fixed learning rate of mean value. Experiments show that our improved algorithm outperforms other traditional methods in real-time performance and stability. If improved method is only used in intelligent visual surveillance system for intrusion detection (intrusion means that the proportion of foreground of the scene is above a constant), the morphology filtering procedure can be ignored because of less effects of noise and slow illumination change, and therefore the computational complexity can be further reduced.After detecting binary foreground of moving objects, we need determine scope, size, location and other information of each object. Because there are noises in binary foreground image, such as holes and isolated points etc, we process binary foreground by using morphology filtering method (opening operation is adopted in this case), and then detect and fill object contours to detect object regions by using cvFindContours and cvDrawContours functions of OpenCV. Finally, we draw frames for object regions. Better experiment results can be achieved.