Distributed Dynamic Task Allocation in Clusters of Embedded Smart Cameras

Traditional surveillance systems of the first and second generation employed analog CCTV cameras to capture the observed scene. The video signals were transmitted to the backend system, where the analog signals have been digitized, analyzed and stored by high-performance workstations. Third generation surveillance systems, however, make use of digital cameras, which compress and transmit the video data digitally. Novel surveillance systems, beyond the third generation, will employ autonomous smart cameras to create distributed surveillance systems, that perform selfmonitoring, and self-reconfiguration during runtime to improve fault-tolerance and the overall quality-of-service. Smart cameras are the key components in these surveillance systems. They not only capture and compress the video stream, but also perform sophisticated, real-time, on-board video analysis of the captured scene. Typical surveillance tasks, implemented by video analysis algorithms, usually include video compression, detection of stationary vehicles, detection of wrong-way drivers, and the determination of traffic statistics like average speed, and vehicle classification. To allow the distributed surveillance system to monitor and reconfigure itself, a methodology is required that assigns and allocates tasks to smart cameras. However, this task allocation methodology has to consider the real-time constraints of the surveillance tasks. This thesis presents a distributed task allocation methodology, that allocates tasks within groups of smart cameras, so called surveillance clusters, dynamically during runtime. In order to guarantee the satisfaction of the tasks’ real-time constraints, the presented task allocation methodology uses the resource utilization of the smart cameras, and the resource requirements of the tasks to find feasible allocations of tasks to smart cameras. By using a cost function, tailored to embedded systems, the optimal allocation is determined in real-time. The thesis deals furthermore with the design of the hard and software architecture for an embedded smart camera. This smart camera (“SmartCam”) is designed for flexibility, high-performance, and scalability, to perform the listed surveillance tasks in real-time, and to support the dynamic allocation of tasks. The scientific contribution of this thesis includes (1) the formal description of the task allocation methodology as a distributed constraint-satisfaction problem (DCSP), (2) the integration of the cost function into the DCSP algorithm to reduce the number of feasible allocations and the running time of the algorithm, (3) the design of a method to reuse previously determined allocations to improve the performance, and (4) the design of the hard and software architecture of the smart camera. Finally, this thesis reports on the prototype implementation of the smart camera and the task allocation system as a case study to prove the feasibility of the approach and to evaluate the task allocation methodology in terms of running times, and the number of allocations.