Dynamic view selection for multi-camera action recognition

摘要

For multi-camera human action recognition methods, there is often a trade-off between classification accuracy and computational efficiency. Methods that generate 3D models or query all of the cameras in the network for each target are often computationally expensive. In this paper, we present an action recognition method that operates in a multi-camera environment, but dynamically selects a single camera at a time. We learn the relative utility of a particular viewpoint compared with switching to a different available camera in the network for future classification. We cast this learning problem as a Markov Decision Process, and incorporate reinforcement learning to estimate the value of the possible view-shifts. On two benchmark multi-camera action recognition datasets, our method outperforms approaches that incorporate all available cameras in both speed and classification accuracy.