Real-time object tracking based on the relative hist model within particle filter framework

Abstract

This paper proposes a novel real-time tracking algorithm by using the relative hist model within particle filter framework. The target-region is roughly enclosed with a rectangle as usual, and color features are used to describe all the types of regions by calculating their color histograms. Inevitably, the background pixels are included in the target-region and the histogram of target-region may be corrupted when performing tracking algorithm. Even the target fails to track. Thus, the relative hist model is proposed to reduce the influence of background pixels. In this model, we not only consider the similarity between the candidate-region and target-region, but also consider the similarity between the candidate-region and background. In other words, the relative hist model tries to find a candidate-region which is more similar to the target-region but less similar to the background. By adopting this model, our tracking algorithm can accurately track the object in real-time. Experiments are performed in various tracking scenes. The experiment results show that our algorithm is of appealing with respect to robustness for real-time object tracking against various backgrounds.