A Novel Moving Object Detection Model Based On Image Registration Within Time Sliding Windows

Moving objects detection and tracking is one of the important research topics in the field of computer vision. In a growing number of applications, such as security surveillance, tracffic monitoring, augmented reality, moving objects detection and tracking is playing a pivotal role. Besides, moving objects detection and tracking is one of the important research topics as a part of technological projects such as vehicle navigation system and robots vsion system and so on.In this paper, we present a novel moving objects detection model named IRTSW-model (a moving objects detection model based on Images Registration within sliding Time Window). In our approach, the video is divided into many small video clips in accordance with time windows. Thus, the prolonged complex camera motion in the entire video is decomposed into short and simple motion in each clip. And in each clip, the movement of camera can be approximately regarded as the composite of self-rotation, small-scale translation and zoom. In this situation, according to the principles of the homography, all frames in the clip can be mapped into a same coordinate space, where the relative movements between the camera and background is eliminated and pixels corresponding to the same point of scene in different frames are mapped into the same location. After then, background model is built by a unsupervised codebook model. At last, pixels of moving objects in each frame are classified using the background model. The unsupervised codebook model is an improved version of the classic one. It determines whether a pixel should belong to the background or a moving object depending on the statistical characteristics of the value sequence at the mapped location. The main contributions can be summarized as follows:(1). In this paper, we use a method of sliding time windows which effectively save computer’s CPU and memory resources to process prolonged video.(2). We propose a novel background modeling method for dynamic scenes. Images registration procedure eliminates the relative movements between background and the camera, and makes it possible to build a background model for a dynamic scene just like for a static scene. And it is also possible to transplant some robust and efficient moving objects detection algorithms in static scenes to the dynamic scenes, which would offer new roads to detecting moving objects in dynamic scenes. (3). A new unsupervised codebook model is presented in this paper. It avoids the training procedure, and classifies pixels depending on the statistical characteristics of the observation value sequence gotten after images registration.