Human Tracking based on Particle Filter with Adaptive Local Descriptor
Sangeun Lee, Keiichi Horio · 2013
This paper presents a human tracking algorithm based on Particle Filter with Local local descriptors in complex environments such that significant occlusions, motion changes and abrupt motion changes occur. The local descriptors are selected for drastically changing geometric appearances over time. To track a target robustly, we suggested Reliable Appearance Model (RAM) that can deal with the complexity of the target appearances in the image sequences. The RAM, a set sparsely selected local descriptors by boosting algorithm in the target image, employs as feature to the observation model in Particle Filter. The likelihood responses in the observation model are calculated between the corresponding feature and observed state in location based on RAM. Since the feature selected by RAM is determined at the first frame in the image sequences, the feature is insufficient to track the target accurately and required to update over time. The architecture separated tracker and feature selector for updating feature is designed. Based on robustness measure, the feature is adaptively modified and redefined to observation model for likelihood responses. The experimental results demonstrate that the present approach could track the target with changing geometric appearance accurately.