Research On Background Extraction Algorithm In Dynamic Scene

This paper presents a dynamic background extraction algorithm which bases onsalience detection. The purpose of us is to apply this algorithm to moving-targetdetection and tracking. There have been many kinds of algorithms aboutmoving-target detection at home and abroad which including frame differentialmethod, optical flow method and background subtraction. Among of them,background subtraction is researched by more and more researchers. The key point ofbackground subtraction is background modeling and real-time update to achievemoving-target detection. In the numerous background modeling algorithm, mostalgorithms are based on single pixel, for instance Single Gauss, Gaussian mixture andAdaptive model. But the Kernel density estimation model is not same as them.Natural scene is full of diversity and uncertainty, generally speaking, which canbe divided into several kind:1) the change of light intensity a nd light direction;2)background objects in sports;3) the same background region with discontinuous orperiodic motion;4) the target in a sudden stop in movement;5) the target shadow,affect the real shape of the target, detection error will appear in the process of targetdetection. The background will be detected as target, target is empty or detectingshadow as the goal, and so on. In order to solve these problems, there happened moreand more algorithms. Each one or more of them generally can solve the problemabout complex scene better. Similarly, kernel density estimation model is differentfrom other model, it has unique advantages. The background modeling process doesnot require parameter selection and update, furthermore, it takes the spatia lcorrelation between pixels into account. Therefore, the moving targets can be detectedperfectly. However, the algorithm is more sensitive to background variation. Part ofthe background with variation will be detected as moving targets.In order to alleviate this problem, this paper presents a dynamic backgroundextraction algorithm which bases on salience detection. The main steps of algorithmincluding: firstly, applying kernel density estimation algorithm to the video sequencefor background extraction, and moving targets detection. Because, there are periodicmotion in the background, therefore the target from background extraction algorithmcontains a large amount of background pixels. At the same time, moving objects in theforeground can be more complete. In order to eliminate the background interferenceof moving target detection which including periodic motion, this paper will applysalience detection method to the kernel density estimation model. Increase the salience region detection on target detection results. Increasing the weight of correctdetection results, reducing the weight of salience region of background, the movingtarget will be detected more accurately. Through a series of experiments, it can beseen in this article the methods can be relatively accurate and complete to detect themoving targets. Experimental results shows that small fragments around the edge ofmoving target is ugly, these small fragments of the pixel is not target, so openoperation has been used to the results. The circular is taken as structural elements, tooptimize the result of target detection which based on salience detection. Therefore,the target edge is more smooth, more clear.