Human Motion Tracking with Weak Prediction

It has been a challenge to capture rapid human motion with self occlusion. Current algorithms are not capable of tracking rapid motions with self occlusion: features with rapid motion are beyond small interest region search, and positions of the occluded features are difficult to be estimated. In this paper, we present a robust human motion tracking algorithm with weak prediction. Instead of predicting the position of each human feature, the region of the whole body is estimated and candidate features are extracted through the overall search in the estimated region. A multi resolution search strategy is proposed to improve the efficiency of overall search: the initial candidate feature set is extracted from the low resolution image and successively refined at higher resolution levels. To establish the correspondence between the candidate and the actual features, an adaptive Bayes Classifier is constructed based on the time varied models of feature attributions, viz. color and motion. And a hierarchical human feature model is adopted to verify and accomplish the feature correspondence. The experiment demonstrates the effectiveness of our algorithm.