This paper addresses the problem of node dynamic selection in camera networks. A selection method based on reinforcement learning is proposed in which the node is selected to maximize the expected reward while minimizing the switching with Q-learning. To accelerate the convergence of Q-learning, the geometry of camera networks is considered for initial Q-values and a Gibbs distribution is used for exploitation. In order to evaluate visual information of the video, a function of the visibility, orientation, definition and switching is designed to assess the immediate reward in Q-learning. Experiments show that the proposed visual evaluation criteria can capture the motion state of the object effectively and the selection method is more accurate on reducing cameras switching compared with the state-of-the art methods.