Automated human activity analysis for multi-camera networks requires algorithms that are both accurate and efficient for practical, real-time use. Current approaches face a trade-off between accuracy and speed, with the most accurate methods having high computational cost. This work is motivated by the observation that multi-camera networks provide redundant data across views. In many cases, algorithms could perform accurate analysis with a subset of the available information. We propose algorithms that dynamically select a subset of the network cameras for each stage of the automated analysis pipeline. The goal of this research is to develop algorithms for multi-camera networks with high computational efficiency without sacrificing accuracy. In particular, we focus on solving core computer vision problems related to the human motion analysis pipeline: detection, tracking, pose estimation, and action recognition. Experiments on benchmark datasets demonstrate the applicability of dynamic view selection to each of these areas.
