Research On Method For Moving Detection In Video Images

Abstract.Moving target detection is an important branch and basis in the field of computer vision. It plays a pivotal role in traffic monitoring, security monitoring, military guidance, visual navigation and video coding, so it has broad prospects and is researched by many scientists, becoming one of the hot research topics in computer vision. In order to identify the target accurately, track the target as well as understand its behavior, it is necessary to detect the moving targets from video sequences accurately. However, because of the complex and diverse environment and the complexity of the real moving target detection, detecting moving targets precisely still faces great challenges. Based on the comprehensive basis of previous researches, the thesis mainly studies the moving target detection under static background.For different monitoring scenarios, moving target detection algorithm is also different. Based on the analysis of traditional algorithms of moving target detection (background subtraction, temporal difference method, optical flow method), we compare background subtraction and temporal difference through experiments. Considering background difference method can detect moving target with high accuracy, but the quality of the background will affect the accuracy of detected targets, we focus on background modeling and updating method. This thesis chooses the multi-frame averaging method with simple principle, Surendra background modeling, single Gaussian background modeling, Gaussian mixture background modeling to extract the background, which were integrated into the correlation of video frames. Then we verify the quality of the background of each way. At the issue of updating Gaussian mixture model, the thesis combine with previous experience, overcome the shortcomings of lack of convergence and low learning rate and use a variety of parameters of the learning rate, making the established background reflect the real background image accurately and timely.This thesis overcomes the shortcomings of temporal difference and background subtraction method, combining the two algorithms. Finally, we program to demonstrate the following five methods to detect target,namely the combination of single-Gaussian and symmetrical frame difference, Gaussian mixture and temporal difference, single Gaussian and multi-frame averaging method to detect objects, symmetrical frame Difference method, temporal difference combined with edge detection. Experimental results show in relatively stable background scenes, each method can better detect the presence or absence of targets, and get to know the target’s shape, size, location and other more comprehensive information, achieving the ideal goal detection.