Robotic Camera Array Motion Planning for Multiple Human Face Tracking Based on Reinforcement Learning

Active camera networks have been widely applied for mobile target sensing. However, traditional model/rule-based methods run into bottlenecks when dealing with complex dynamic scenarios due to their lack of long-term planning abilities. This article presents a mobile camera array consisting of multiple robotic cameras, which actively arrange their locations and poses to increase the human face tracking efficiency by leveraging reinforcement learning (RL) techniques. The novelty of the proposed system is threefold: 1) developing a distributed RL network based on the soft actor-critic (SAC) model to increase the learning efficiency and to capture targets' long-term behaviors; 2) developing a reward function for the RL networks to facilitate the cooperation among multiple cameras; and 3) developing the self-attention modules in the distributed RL networks to enhance the multiagent global state estimation from local cameras. The developed system is validated through human face detection and recognition applications. The simulation and real-world experiment results demonstrate the efficiency and superior performance of the proposed methods.