Infrared target tracking using multiple instance learning with adaptive motion prediction and spatially template weighting

In this paper, we formulate the problem of infrared target tracking as a binary classification task and extend the online multiple instance learning tracker (MILTracker) for the task. Compared with many color or texture based tracking algorithms, the MILtracker highlights the difference between the target and the background or similar objects, and is thus suitable for infrared target tracking which undergoes serious textual information loss. To address the specific challenges in the infrared sequences, we extend the original MILtracker from two aspects. Firstly, an adaptive motion prediction procedure is integrated in to enhance the efficiency of the tracker. This step helps discriminate disturbing objects that are visual very similar to the target under tracking. Secondly, a spatial weight mask is introduced into the target representation to augment its robustness against similar background clutters, especially distracters. We apply the proposed approach on several challenging IR sequences. The experimental results clearly validate the effectiveness of our method with encouraging performances.