In previous work, we proposed a human tracking algorithm based on the reliable appearance model (RAM). The RAM is a set of discriminative local image descriptors that is selected by a boosting algorithm to identify a target in the initial frame, and is employed as an observation model in a particle filter. As the appearance model of the target in human tracking constantly changes as time passes owing to changes in pose, it is necessary to adaptively update the RAM to improve the tracking accuracy. In this paper, if necessary, an insufficient local image descriptor for robust tracking is updated. In order to classify whether local image descriptors are suitable or not during tracking, a distance histogram corresponding to a local image descriptor is constructed. When the histogram indicates that the local image descriptor is lacking in tracking performance, then it is updated. The experimental results demonstrate that the adaptive appearance model successfully tracks sport players even when their pose often changes.
展开▼
