Rights Creative Commons: Attribution 3.0 Hong Kong License Arm Pose Modeling for Visual Surveillance

Abstract—Arm pose of individual person often indicates his/her behavior and interaction in a crowd. This paper focuses on how to model arm pose with respect to walking styles in visual surveillance. We propose to use a Bayesian network to represent the relationship between the upper arm, forearm and hand. Two sets of probabilistic templates are used to estimate the likelihoods of these three arm parts against their boundary, foreground and skin color features. The prior distributions, a critical part of the Bayesian formulation, are estimated using a kernel density estimator based on the CMU motion dataset, whereby MAP is applied to estimate the part’s orientation and size before fitting them to the arms. The proposed method is found to perform effectively on the CAVIAR sub-dataset and a HKU dataset, and compare well against another popular method. Potentially, it can be extended to track arm movements or to analyze the objects carried by individuals