Dynamic node selection in camera networks based on approximate reinforcement learning

In camera networks, dynamic node selection is an effective technique that enables video stream transmission with constrained network bandwidth, more economical node cooperation for nodes with constrained power supplies, and optimal use of a limited number of display terminals, particularly for applications that need to obtain high-quality video of specific targets. However, the nearest camera in a network cannot be identified by directional measurements alone. Furthermore, errors are introduced into computer vision algorithms by complex background, illumination, and other factors, causing unstable and jittery processing results. Consequently, in selecting camera network nodes, two issues must be addressed: First, a dynamic selection mechanism that can choose the most appropriate node is needed. Second, metrics to evaluate the visual information in a video stream must be modeled and adapted to various camera parameters, backgrounds, and scenes. This paper proposes a node selection method based on approximate reinforcement learning in which nodes are selected to obtain the maximum expected reward using approximate Q-learning. The Q-function is approximated by a Gaussian Mixture Model with parameters that are sequentially updated by a mini-batch stepwise Expectation–Maximization algorithm. To determine the most informative camera node dynamically, the immediate reward in Q-learning integrates the visibility, orientation, and image clarity of the object in view. Experimental results show that the proposed visual evaluation metrics can effectively capture the motion state of objects, and that the selection method reduces camera switching and related errors compared with state-of-the art methods.