Comparative analysis of methods for video tracking algorithms improvement

In this paper we made a comparative analysis of methods (based on Lipschitz index, Lyapu-nov index and Forward - Backward criterion) for evaluation of video tracking algorithms (Mean-Shift, Particle filter). Developing tracker - a software and hardware system for object tracking in video sequence, is one of major tasks in computer vision. Mathematically, object is modeled by set of numerical characteristics - feature vector, which allows detecting object on a video frame and then tracking this object on further times. We investigate object lost in video sequence, due to occlusions with another objects and/or obstacles; similarity by feature vector of tracking object and background and etc. As criteria of tracker reliability we suggest to use threshold conditions when comparing object's feature vector in tracking algorithms. Let's call external Forward - Backward criterion and others, based on consideration tracker as dynamic system (Lipschitz index, Lyapunov index). Quantitative estimates of tracking algorithms evaluations are obtained. These allow using suggested indices for combining different tracking algorithms and feature vectors in data fusion frameworks. This research is done in Tomsklabs PTE LTD.

Keywords

Authors

References