A majority of tracking algorithms have a number of parameters that require tuning and often significantly influence tracking performance. Without a formal procedure of finding the best parameter values, the authors usually set them manually by trial and error. In this thesis we focus on development of a system for systematical analysis of trackers parameters. We propose an iterative approach to solving the parameter optimization task. With steepest ascent hill climbing search method, we narrow the search space in each iteration and move in a direction of the best solution. The optimization is based on a recent tracker evaluation measure which considers tracker reinitialization and represents the expected average overlap. We implement the proposed methods by expanding an existing tracking evaluation system. We propose procedures for result visualization, which enable a detailed analysis and insight into parameter properties. We also propose a procedure for determining how successful the optimization method is. As a proof of concept, we test our system and proposed methods on a Mean Shift tracker. By comparing results obtained by our optimization approach, we conclude that we can improve tracker performance by improving the parameter values. Experimental results indicate approximately 5% improvements over original published parameters.
|
