Motion Detection And Tracking Based On Gaussian Mixture Model And Kalman Filter

Moving object detection and tracking is an active topic in video processing. With the development of artificial intelligence and information technology, moving object detection and tracking technology has already been applied in military, industry, human-computer interaction, intelligent transform and science researching, etc. After deep study on present algorithms and existing problems of moving object detection and tracking technology, the thesis is formed to improve the accuracy and robustness of object detection and tracking, a new system based on Gaussian Mixture Model and Kalman filter was proposed. The main contributions are as follows:Firstly, two methods of moving object detection, optical flow calculation and symmetrical differencing, has been researched systematically. Though calculation optical flow costs high computational complexity, this method can detect moving target without any other prior knowledge. Optical flow implies not only moving information about target, but also abundant information of texture. So it is a good supplement of other methods. This thesis combined it with symmetrical differencing to detect moving objects. Furthermore, experimental analysis was presented.Secondly, Gaussian Mixture Model can adapt to variance of complex background, and it is a real-time algorithm of moving object detection. However, the detection result exist much noise and holes. The thesis improves Gaussian Mixture Model through adding background subtraction to present Gaussian Mixture Model algorithm, and the holes inside target are filled up by determining the location of contours. Experimental results show that our method can decrease noise and fill up holes inside target.At last, as to moving object tracking, connected component fields defined as blobs were found in the result image of our detection stage. Blobs'information such as position, size and spatial moments are calculated in order to match corresponding blobs. Kalman filter is applied to estimate position of blobs. Experiments have shown effective and fast-processing result.