An Enhanced Model-based Tracking Algorithm with Dynamic Adjustment of Learning Parameters according to Online Performance Evaluation
Yehong Chen, Pil Seong Park, Qian Gao · Indian Journal of Science and Technology · 2015
In online object tracking, the ability to identify possible track loss without ground truths is important. Template Inverse Matching (TIM) is one such performance evaluation method that does not require ground truths, but it is only applicable to template-based methods. We extend the idea of TIM to a more general method so that it is applicable to model-based trackers. The idea can be introduce to any model-based trackers to adjust model learning parameters whenever possible track loss is identified by this technique. As an application, we introduced the idea to the famous Real-time Compressive Tracking (CT for short) algorithm and the results are compared with that of the original CT. The experimental results show better performance and stability with negligibly small additional running time. Keywords: Model Inverse Matching, Object Tracking, Online Performance Evaluation, Real-Time Compressive Tracking, Template Inverse Matching