Modeling and optimizing the coverage of multi-camera systems

Aaron Mavrinac, University of Windsor


This thesis approaches the problem of modeling a multi-camera system's performance from system and task parameters by describing the relationship in terms of coverage. This interface allows a substantial separation of the two concerns: the ability of the system to obtain data from the space of possible stimuli, according to task requirements, and the description of the set of stimuli required for the task. The conjecture is that for any particular system, it is in principle possible to develop such a model with ideal prediction of performance. Accordingly, a generalized structure and tool set is built around the core mathematical definitions of task-oriented coverage, without tying it to any particular model. A family of problems related to coverage in the context of multi-camera systems is identified and described. A comprehensive survey of the state of the art in approaching such problems concludes that by coupling the representation of coverage to narrow problem cases and applications, and by attempting to simplify the models to fit optimization techniques, both the generality and the fidelity of the models are reduced. It is noted that models exhibiting practical levels of fidelity are well beyond the point where only metaheuristic optimization techniques are applicable. Armed with these observations and a promising set of ideas from surveyed sources, a new high-fidelity model for multi-camera vision based on the general coverage framework is presented. This model is intended to be more general in scope than previous work, and despite the complexity introduced by the multiple criteria required for fidelity, it conforms to the framework and is thus tractable for certain optimization approaches. Furthermore, it is readily extended to different types of vision systems. This thesis substantiates all of these claims. The model's fidelity and generality is validated and compared to some of the more advanced models from the literature. Three of the aforementioned coverage problems are then approached in application cases using the model. In one case, a bistatic variant of the sensing modality is used, requiring a modification of the model; the compatibility of this modification, both conceptually and mathematically, illustrates the generality of the framework.