A new unified framework for object detection and tracking in infrared imagery

Abstract

In this paper, we propose a novel general framework for target detection and tracking in infrared image sequences. An integrated tracking system is described by this framework based on multiple models learning online. The relations among each component of the tracking system are expressed distinctly. Furthermore, we emphasize that the main components of the tracking system shouldn't be invariable. On the contrary, they should update dynamically. An integrated tracking system is composed of six modules. The target appearance will change as the target object moves from one place to another. So the object description also needs update dynamically in the tracking framework. At the core of many approaches for object tracking is the metric or similarity measure used to determine the distance between the target template and candidates. In the proposed tracking framework, the distance measure is learnt online and update dynamically by the ensemble learning algorithm. Approaches on estimation of object tracking can be divided into two groups: deterministic approaches and stochastic approaches. In our unified framework, the estimation approach is not fixed, but adaptive. The observation model, motion model and number of particles can adapt to the changes of the foreground and background. Our extensive experiments show that the presented algorithm performs robustly in a large variety of infrared image sequences. The approach proposed in this paper has the potential to solve other sensor fusion problems.