Research On Video Object Tracking Technology Based On Mean-Shift Algorithm And Kalman Filter

Nowadays, video target tracking has been a hot subject in the field of computer vision research, and has been widely applied in video surveillance, visual navigation of robots, military guidance, medical diagnose, and human-computer interaction, etc. Video target tracking refers to detecting, extracting, identifying and tracking for the moving object in the video sequence, thus providing foundation for further video analysis and understanding. In the real world, it is difficult for some frequently occurring instances, a few among which are the variation of the state of the moving targets, the occlusion, object deformation, illumination variation, background noise, etc. Although a lot of scholars have made research progresses in video target tracking techniques, there still exist many unsolved difficulties in tracking the targets for a great number of complex environments.This thesis firstly presents a review and analysis of the most updated technologies on video moving object detection, with focus on intra-frame subtraction and background updating, followed by mainly investigating the tracking quality using the Mean Shift algorithm and the Kalman filter based on the target detection. The research efforts are mainly given to improve those existing tracking algorithms for adopting them to the real-world applications.1. To solve the tracking problems of target transformation, partial occlusion and over-fast motion identified with objects, this thesis proposes an alternative approach using self-adaptive Mean-Shift algorithm combined with Kalman filtering techniques for efficient tracking of the video targets. Firstly, in the initial frame the target is determined, followed by computing its H component in the histogram, then the target within the window of the frame in question is transformed to the image of probability. In current frame, the Kalman filter is adopted to predict the very initial position of target, and the self-adaptive Mean-Shift algorithm is to be applied to detecting the video target. The proposed method is put to the tests, which suggest satisfactory tracking results for both rigid and non-rigid objects, as well as multi-objects, with good self-adaptability.2. This thesis also brings forward a video object tracking method combining the Bhattacharyya coefficients maximization with spatio-temporal information. The Kalman filter is used to predict the target movement information in the time domain, while in the space domain the target is precisely matched through using Camshift algorithm. Due to the strong maneuverability of the moving target, there exists a relatively big discrepancy between the predicted and true position, which will cause failure in tracking for the upcoming frame. To deal with this problem, the thesis adopts a kernel matching approach based on Bhattacharyya coefficients in a rough to precise way. The size of self-adaptive search window is properly increased based on the position of the prediction window, and the initial matching window is defined according to the Bhattacharyya coefficients maximization. Finally, the target is precisely matched by applying Camshift algorithm. The tests are conducted, suggesting that the proposed method is highly precise in tracking fast and maneuvering moving target.