Research On Method Of Detecting Moving Target Based On Omnidirectional Vision

Detection of moving objects includes video or image sequence acquisition, image processing, determination of target location, target identification and target tracking. Moving target detection requires separate foreground from background to detect image motion and prepare for target tracking and trajectory analysis. Therefore, detect moving objects accurately is essential technologies, and also the premise of following works.Panoramic vision system could observe 360°scene simultaneously. Catadioptric panoramic vision system is consist of a mirror and a camera. The panoramic image can easily transform the input image to human friendly images. These advantages are suitable for many applications that requiring a large field of view. In this paper, we research on moving target detection based on panoramic vision system for this system is able to obtain a large field of view.First of all, we tried to realize target detection based on methods of template matching. According to the different measurement methods of correlation, the calculations of correlation functions are also different. In this paper, we study general template matching methods and sequential similarity detection algorithm (SSDA). We obtained template matching results through experiments and compared the length of time of different image template matching costs.Secondly, we study the morphology-based image motion detection algorithm. Moving objects are detected from difference of images. The noise caused by image difference was removed by binary morphology. The moving object was tracked by the centroid detection method. Experimental results show that the method based on morphology can be used as an effective target detection and tracking techniquesFinally, we focused on Lucas-Kanade optical flow method for motion detection. Methods of Harris corner detection and sub-pixel corner detection were discussed, followed by studies of the Lucas-Kanade algorithm, then the definition of image pyramid algorithm was discussed which could make up the deficiencies of LK optical flow algorithm. At last, the vector field of the motion was obtained through the experiment.