The focus of this thesis is on dynamic clustering in resource-aware visual sensor networks (VSNs). A VSN consists of spatially distributed visual sensor nodes. They have the ability to retrieve the observations from the environment by an image sensor, process these observations locally and exchange the pre-processed data among the other nodes in the VSN. These nodes are usually battery-powered with limited resources in processing, storage and communication capacity. Furthermore, they have overlapping and non-overlapping field of views (FOVs). The overlapping case is preferable as it enables coordination and control among the nodes as simultaneous observations of a specific object are provided. Coordination of the nodes is referred to the data exchange and processing in a VSN controlled by the local observations and resources. These observations allow the cameras to exchange and optimize their local object's state, typically presented as position and velocity. There are many different ways to design and deploy a VSN with its nodes. Therefore, a design space is explored to show the most important design criteria in four dimensions describing the design possibilities of the node and protocols in the VSN. The design space assigns, e.g., a fully distributed state estimation on the object's observations. Therein, each node exchanges its states with all other nodes in the network. This wastes a lot of resources in processing, storage and communication and introduces redundant data on all nodes. Contrary to the distributed, a dynamic clustering approach is used to coordinate and control the visual sensor nodes of a VSN. This protocol is called Resource-Aware dynamic CLustering (oRACLe). A visual sensor node can become a cluster member only if it has the same object in its FOV. Moreover, there is only a single visual sensor node the cluster head that gets the responsibility of processing computational complex tasks, as given with state estimation. For the composition of a cluster, a market-based approach is used, where after an auction initiation of the cluster head, the cluster members can bid for the specific object, if it is in their FOV. This bid is constituted with the local tracking performance on the object and the locally available resources. For evaluating the proposed protocol a new simulator for VSNs is designed and deployed. The simulation results show that with the dynamic clustering utilized with a market-based approach, the resources can be decreased drastically when comparing the proposed protocol with the fully distributed approach.
|