Target Container: A Target-Centric Parallel Programming Abstraction for Video-based Surveillance
MetadataShow full item record
Surveillance systems are some of the most computationally intensive applications. Despite technological advances, lowcost of sensors, and continuous improvement of computer vision algorithms for analyzing video data, large-scale reliable surveillance systems are yet to become common. We argue that building effective smart surveillance systems requires a new approach, with a focus on programmability and scalability. We define programmability as the productivity of building concurrent and parallel systems. While considerable progress has been made in the area of computer vision algorithms, such advances cannot translate to deployment in the large until adequate system abstractions and resource management techniques are in place to ensure their performance. In this paper we propose a novel abstraction, the target container (TC). A TC is a programming and execution abstraction that allows programmers to focus on targets, simplifying the programming effort while allowing better resource utilization under overloaded scenarios. A TC provides a single interface for video surveillance application programmers to operate on all video camera image streams for a particular target. Each TC is mapped to a single target and aggregates the system resources dedicated to that target. By mapping all of the individual sensor streams presently containing data for a particular target, the TC system is able to more efficiently manage per-target resources and provide an execution model that is target-aware, and improves the overall scalability of a surveillance system in terms of the number of targets the system can track in real time. The TC model enables a variety of target tracking policies, including target prioritization. Under this policy, the TC system automatically allocates system resources based upon programmer-specified target priority, enabling a TC-based surveillance system to guarantee that high-priority targets are tracked, even under conditions of system overload (i.e., when many targets are being tracked).We have implemented an experimental prototype of the TC model, and measurements confirm its performance benefits.