Application Of View Morphing On Continual Visual Object Tracking

In this thesis,single target visual tracking system was studied,and the algorithm was extended to the situation of continual tracking via multiple cameras.By introducing the improved view morphing into the tracking system,this thesis implemented two typical discriminative tracking algorithms.Then,by integrating the object detection and re-identification module,the tracking system could run smoothly under multi-camera situations.Finally,we used several open datasets with self-designed indicators to verify the improvement of system.Experiments showed that the precision and robustness of short-term tracking and object detection were improved after adding view morphing module.During the tracking process,the tracked object is often constantly changing.The rotation in plane may invalidate some non-rotational invariant features and lead to model drift and tracking failure.Due to the out-of-plane rotation like perspective transformation,it is difficult to consistently and stably express the target.In this thesis,we used and improved the view morphing module to measure the constraints among the parts of target,speculated the spatial structure of the target and generated multi-view projection images as addtional samples,and improved the trackers' classifier performance.In multi-camera tracking scenarios,the object is moving between the fields of view of different cameras.Initial frame should be given by computer for each camera to track.In this thesis,target detection and re-identification technology were introduced to build multi-camera continual tracking system framework.The view morphing module was also applied to improve the performance of detector.At the same time,a target detection and recognition metric during tracking was proposed.Through this metric,enhancement of the classifier by improved view morphing was verified.Improved view morphing enhances the generalization ability of classifiers by providing multi-angle samples,and ultimately improves the accuracy of classifiers.In the future,view morphing could be applied in more applications with limited number of samples and constrained diversity as a kind of data augmentation method.