Dynamic Camera Network Reconfiguration for Crowd Surveillance

Dynamic Camera Reconfiguration with Reinforcement Learning and Stochastic Methods for Crowd Surveillance

“…Table 1 summarises the results obtained using our greedy approach [ 9 ]. In experiments (3) and (6), is set to 1, causing our camera network to focus only on observing pedestrians with no incentive to explore new areas in the environment.…”

“…The policy presented in [ 9 ] and reported in Section 3.3 suffers from two main limitations: The reconfiguration objectives are the same for the different camera types, namely UAVs and PTZs. In the real world, UAVs have a higher cost of deployment and movement with respect to PTZs, while they provide more degrees of freedom for their reconfigurability.…”

“…Leveraging on simulators and virtual environments can be an effective tool to deal with these limitations. Virtualisation paradigms have been exploited both in camera surveillance [ 5 , 6 ] and crowd analysis [ 9 , 20 ].…”

“…Ensuring both coverage of the whole monitoring area and a good video quality of moving individuals is challenging using non-reconfigurable (fixed) cameras [ 5 , 9 ]. An high number of fixed cameras would provide the required coverage of the scene, but at a high cost.…”

“…In [ 9 ], we proposed a greedy approach to control the trade-off between covering the widest possible area of the area of interest ( global coverage ) and focusing on the most crowded parts of the scene ( people coverage ). In our previous work, we proposed a decentralised greedy empirical approach, where each camera aims at optimising the coverage performance in its local neighbourhood.…”