Human tracking with thermal omnidirectional vision

In this paper, we explore a new tracking system for human tracking in thermal catadioptric omnidirectional vision. Due to very limited features can be adopted in thermal image except for contour information, we proposed to use Histogram of Oriented Gradient (HOG) feature to represent the contour information and employ Support Vector Machine (SVM) to classify the foreground and background. In this paper, there are three novel points. First, the classification posterior probability of SVM will be adopted to relate the observation likelihood of particle filter to guide the particles for tracking purpose instead of neglect in previous tracking method. Second, due to no existing thermal catadioptric omnidirectional vision database available in public, a thermal catadioptric omnidirectional video database and extracted human samples have been established for academic studies. Third, tracking window distribution of particle filter has been adjusted to fit the characteristic of catadioptric omnidirectional vision on account of the size of target in image is varying when the distance between target and omni-sensor changed in world coordinate. In addition, the catadioptric omnidirectional imaging is different with traditional imaging for inherent distortion, so the polar coordinate will be used. The experimental results show that the proposed tracking approach has a stable performance.