Research On Dynamic Node Selection In Camera Networks

Camera networks have recently gained significant interest from research point of view for the effective resolution of partial and full occlusions and continuous tracking of targets over large areas where the limited field of view (FOV) of a single camera is insufficient. Such networks are extensively used in security, surveillance, human-computer interaction, navigation, and positioning fields. Increasing the size of the network camera and the number of nodes is challenging for network node management and control; dynamic node selection in camera networks has particularly attracted attention because of its ability to tradeoff between the requirement of the vision task and constraints such as network transmission, computing power, and limited number of display terminals.In this thesis, we studied the problem of dynamic node selection in camera networks for intelligent surveillance and specific target tracking. We researched on several techniques of state modeling and planning in an uncertain environment, online learning, and the optimization of selection policy and its approximate representation.The main contributions of this paper include the following three aspects:1) We proposed a state modeling and planning method in dynamic node selection in an uncertain environment for intelligent surveillance.For system state modeling, the camera visual score and the previous selection result are modeled as a system state based on a partially observable Markov decision process model, and the belief states of the model are used to represent noisy visual information; state transition was defined with it. At the same time, we use the visual score from each camera as a positive reward and camera switching cost as a negative reward to measure the degree of optimization resulting from that action. A point-based approximation method was used to obtain the solution of POMDP. For visual information measure, an innovative evaluation function identifies the most informative of several multi-view video streams for intelligent surveillance by extracting and scoring features related to global motion, attributes of moving objects, and special events, such as the appearance of new objects. Our proposed POMDP-based method has a higher degree of accuracy and is more stable by reducing frequent switching with noise and unstable visual computing, and the visual score function can accurately reflect and describe the visual information in a scene.2) We proposed a reinforcement learning method of dynamic node selection policy in camera networks for specific target tracking.A reinforcement learning framework to select the most appropriate node dynamically was proposed on the basis of on Q-learning. In this framework, the selection policy was represented as discrete Q-function referring to the expected long-term rewards of selection action, and it was learned and updated with the temporal difference principle by the immediate reward from the selection process. To accelerate the learning, the target’s current state and scene topology were used as a priori knowledge to initialize the Q-function, and an improved exploratory action selection policy using a Gibbs distribution was adopted. For the evaluation of selection, the immediate reward in Q-learning integrated the visibility, orientation, and image clarity of the object in view to determine the most informative camera node dynamically for specific target tracking. Our proposed method of policy learning and updating can reflect the changes of the system effectively and improve the ability to adapt for policy, and the proposed visual evaluation metrics can effectively capture the motion state of objects.3) We proposed an approximate representation method of node selection policy.An approximate Q-learning method of node selection policy was proposed, in which, the Q-function was approximated by using a Gaussian mixture model, and the value of the Q-function was calculated by using a Gaussian mixture regression. To obtain the value of policy while being learned, the parameters of GMM sequentially updated by a mini-batch stepwise expectation-maximization algorithm and the number of Gaussian components were adjusted with posterior probability of a Gaussian component for the sample. The approximate function can represent the node selection policy effectively and has a faster convergence for learning online of policy, and it is better than the discrete Q-function method in terms of the average visual score and switching number.