In randomly deployed visual wireless sensor networks for surveillance applications, the scheduling of sensor nodes can be seen from the risk perspective: different parts of the area of interest may have different risk levels according to the pattern of observed events such as the number of detected intrusions. In this paper, we propose a multiple-level activity model that uses behavior functions to define application classes and allows for adaptive scheduling based on the application criticality and on the availability of multiple cover sets per sensor node. The paper then describes how an adaptive scheduling model can be defined in order to dynamically schedule nodes by varying the capture speed according to nodes' environment. Simulation results are presented to validate the performance of the proposed approach in terms of percentage of active nodes, percentage of coverage and stealth time under intrusion scenarios.