Application And Research On Human Motion Tracking In Large Scene By Using A Panning Camera

Computer vision becomes one of the most popular topics, however, there is nostandard formulation of how the so-called"computer vision problem"should be solved,these problems mainly come from the practical applications, in general, the methods areproofed to be valid for a specific task, but it is hard to make them work well over a widerange of other applications.This dissertation focuses on the key technology of human motion tracking in a largescene by using a panning camera, which include the following aspects: For a large scene,a single panning camera on a tripod has to be chosen, at the same time, zooming is aban-doned. The camera own panning motion needs to be compensated in tracking process byusing the computation of homography of two consecutive frames, it is a random process,therefore, when this computation has been introduced into the tracking process, the com-putation stability of this homography affects the tracking results directly, how to evaluatethe relation of them. At the same time, how to calculate the other homography mappingeach frame into the real scene model fast and accurately, and how to evaluate the preci-sion of this homography. Human motion tracking becomes a challenging task because oflower image resolution, complex background, fast object motion and occlusions, there-fore, more prior knowledge, more accurate object representation and more robust trackerare badly needed.In view of the above problems in the practical application, this dissertation makesuse of the skater tracking in a large rink as an application background, the main researchcontents and innovations include the following aspects:(1) A novel application algorithm framework of human motion tracking in a largescene by using a panning camera is proposed, especially for complex sports applications,it maybe have better robustness and stability.(2) A novel efficient algorithm to reduce the accumulative registration error for along image sequence is proposed, in which any frame are transformed to the real rinkmodel only needs three steps, namely, mapping each frame to its corresponding refer-ence frame, then mapping reference frame to the panorama of the rink, last, mappingthe panorama to the real rink model. Compared with the traditional methods, the pro- posed algorithm avoids to a concatenation multiplication of the homography of consecu-tive frames, therefore, it can reduce the accumulative registration error efficiently and isvery important to improve the stability of the practical application.(3) The hierarchical model, which represents tracked object more accurately, is pro-posed. It can help the tracker to know when and how to update the hierarchical model.That ensures the accuracy of the model update in tracking process.(4) Using multiple cues in tracking process, namely, the template matching methodfor helmet and the color histogram matching for body, that can make the tracker morerobust when a skater moving through the advertisement board or occlusions appear on thecurve.(5) A new method of constructing a color kernel histogram is proposed, it introducesa mask function into the kernel, which can filter the rink pixels considered as a kind ofbackground interference and improve the accuracy of histogram matching efficiently.(6) To make use of UKF (Unscented Kalman Filter) own properties, UKF can cap-ture the mean and covariance of the statistical variables accurately, compared with theparticle filter, it is specified using a minimal set of deterministically chosen sample points,therefore, its computational efficiency is more higher. At the same time, object represen-tation and multiple cues are integrated cleverly into the tracking framework of UKF.(7) This dissertation proposes a novel evaluation of human motion tracking in alarge scene by using a panning camera, which includes two aspects: 1. Compared withthe traditional tracking method, the calculation of the homography of the consecutiveframes, which is used to remove the camera motion, is introduced into the prediction stepof the tracking process. The effect of the introduction of this homography is simulatedby adding many different groups of the noise to the object position marked manually,through which the analysis of the relation between the accuracy of this homography andthe tracking performance is discussed in detail. 2. The accuracy of the homography thattransforms each frame to the real rink is evaluated by an indirect approach, that is to say,the accuracy of this homography can be estimated by the registration error of the markedblock in the coordinate of the real rink. In addition, the factors, which can affect thisaccuracy, have been analyzed in theory.A novel application algorithm framework of human motion tracking in a large sceneby using a panning camera is proposed, it can be extended to other similar applications easily, such as track and field events, cycling and ball games, especially suitable for sportsinformation access in real-time, data statistics and presentation of tactical simulation insports television. At the same time, some of the key technology detailed in this disserta-tion can also be used in robot vision navigation and localization, military, virtual reality,safety monitoring and so on.