The Research Of Integration Of Foreground And Background Cues For Change Detection In Non-static Camera

As the first step of applications of computer vision,change detection has been a challenging problem.It has a wide range of applicants in video surveillance and often be used as the input of other applications.Traditionally,change detection assume a static camera.However,with the situation that large numbers of non-static camera are shown recent years,the assumption of static camera hardly limits the applicants of change detection.And a new challenge-“change detection in non-static camera” is proposed in computer vision.In the videos obtained by non-static camera,the camera motion is mixed with the motion of moving objects.Therefore,the position of any objects could be changed due to the motion of camera.The efficiency of previous work is limited by the motion estimation by which they eliminate the effect of camera motion.Unfortunately,since the complexity of natural senses and the unpredictability of camera motion,it is so hard to produce a perfect motion estimation.To solve this problem,this paper proposed the IFB(Integration of Foreground and Background cues)model.Unlike previous work,IFB model focus on detecting moving objects by integration of rough foreground and background cues instead of improving the accuracy of motion estimation.The main work of this paper is shown as follows:(1)This paper describes the means and the current situation of change detection.It reviews the advantage and the disadvantage of current change detection algorithms,and explains the challenges of change detection.(2)This paper extract the background by homography transformation.It introduces the process of homography transformation and explain who the homography transformation can be used to extract background.Moreover,s series demonstration experiments are used to demonstrate the efficiency of background extraction.(3)This paper proposed the IFB(Integration of Foreground and Background cues)model.In particular,the foreground cues are captured from the extension of algorithms based on static camera,whereas the background cues are extracted from the spatio-temporal features with motion estimation.Then the super-pixels captured under multiple scales are used as the container for integrating foreground and background cues.Benefit from the alternative between the foreground and background cues,the defects of each other are compensated.Therefore,the rough cues are enough for accurately foreground segmentation.(4)Comprehensive evaluations on several standard benchmarks demonstrate the superiority of our work against the state-of-the-art.