Object Tracking Based on Modified TLD Framework Using Compressive Sensing Features
Résumé
Visual object tracking is widely researched but still challenging as both accuracy and efficiency must be considered in a single system. CT tracker can achieve a good real-time performance but is not very robust to fast movements. TLD framework has the ability to re-initialize object but can't handle rotation and runs with low efficiency. In this paper, we propose a tracking algorithm combining the CT into TLD framework to overcome the disadvantages of each other. With the scale information obtained by an optical-flow tracker, we select samples for detector and use the detection result to correct the optical-flow tracker. The features are extracted using compressive sensing to improve the processing speed. The classifier parameters are updated by online learning. Considering the situation of continuous loss of object, a sliding window searching is also employed. Experiment results show that our proposed method achieves good performances in both precision and speed.