Active Vision for Human Action Sensing

This paper presents a real-world implementation of a multi-camera active vision system for form/action recognition of time-varying geometry objects (specifically, the human form). The proposed method uses a novel sensing-system reconfiguration methodology to improve recognition performance. Near-optimal viewpoints are selected by the algorithm, along a prediction time horizon, to maximize target visibility in the presence of multiple, potentially moving obstacles. The proposed reconfiguration method has been tested using controlled experiments.