Optical Flow Based Object Motion Tracking With Cascaded Outlier Rejection

Abstract

Object tracking is one of the main areas in computer vision. Tracking is basically a process of locating a moving object over time using any kind of sensing device. Here the image sensing device works as the input device and the whole tracking is done by processing the image information. To track an object is not an easy task because of various limitations of the input device and the processing ability of the processors. The underlying scientific phenomenon of the dynamics and physics of this mysterious world make it worse. Our tracker was developed to track an object regardless of its any feature, for example, the shape, color or intensity, illumination change. It was also developed to work in fast camera motion movements. To have an idea of the motions of every pixel in two consecutive frames we have to determine the motion vectors at first. Using Lucas-Kanade method we managed to find the motion vectors. Now it is the problem of determining the actual motion vector which has caused the tracked object move. We used the Global motion estimation from the coarsely sampled motion vector field. We incorporated the cascaded outliers rejection method where outliers indicates the noises. It was incorporated after getting the motion vectors at first stage and before calculating the global motion estimation. This system was experimented on three different videos which had distinctive characteristics. Comparing with the state of the art trackers our tracker showed a very good performance and sometimes better than those.