Multiple cues fusion for object tracking with particle filter in infrared image

Abstract

Object tracking in infrared image sequences is a challenging research topic due to the extremely low signal to noise ratio of IR image. In this paper, a new tracking method based on multiple cues fusion particle filter framework is proposed. In order to make full use of the object appearance information, both the spatial distribution and the gray distribution of the object are considered in object modeling. Meanwhile, an affine transform model is used to estimate the motion of the object which is integrated in the tracking framework. Firstly, the motion information is used to represent the state of each particle. Secondly, each object is modeled by intensity template and gray histogram which are independent to each other. The weights of the particles are obtained through the similarity of each feature model. Finally, to overcome the problems relating to the changes in the object appearance, the object model is dynamically updated according to the tracking result using kernel density estimation. It uses the complementarities of the two features to improve the reliability in tracking task. The experimental results show that the fusion of multiple cues makes the tracking performance effective in infrared images.