Moving Object Detection And Tracking In Dynamic Scene

Moving object detection and tracking in dynamic scene has a very wide range of applications. It plays a key role in many fields such as smart video surveillance, traffic monitoring and some military occasions. Background modeling, moving objects detection and tracking is the focus of this paper. There are many experts and scholars have been studied in these fields for many years, and made great progress. However, moving object extraction and tracking still face many challenges in a complex scene with shadows interference and illumination changes. Based on the work of previous people, this paper studies the following aspects:(1) Motion detection algorithm based on background modeling under static imaging platform. We analyze Gaussian mixture model algorithm proposed by Stauffer etc. and give out a set of models update equation to improve the convergence rate of the background model in busy scenes. The learning rate of each model component can adaptively adjust according to its matching degree to the background. Our strategy improves the learning accuracy of the background model. In the outdoor scenes, shadows often interfere with the targets segmentation accuracy and this will affect the follow-up targets tracking and recognition. In this paper, the improved Gaussian mixture model and RGB color space shadow suppression algorithm are integrated to improve extraction accuracy of targets.(2) Object detection under moving imaging platform based on automatic image registration. We mainly study SIFT algorithm and cooperate it with Random Sample Consensus algorithm to estimate transformation matrix between images. After image registration, we use frame difference for moving objects detection. The results show that the algorithm can effectively detect the targets under platform with movement.(3) Particle filters algorithm and multi-object tracking technique. In order to reduce our system complexity and enhance real-time tracking, detected objects in the system are represented by their bounding rectangles. Association matrix is built between two frames for targets association according to a simple kind of cost function of features.(4) Video surveillance software design and implementation. In the final part of this paper we implement a video surveillance software platform which integrates a variety of detection algorithms.