Architectural Considerations for Video Content Analysis in Urban Surveillance

Abstract

Successfully deploying Video Content Analysis (VCA) solutions for urban surveillance poses significant challenges for manufacturers and system integrators. Urban surveillance is typically characterized by a very large number of cameras (thousands and more) distributed over a large area and installed in both outdoor and indoor views. From the user perspective the primary rule of the surveillance system is to provide quick, reliable and high quality access to live and recorded video streams from all cameras. VCA is considered an important but secondary functionality that is required in order to provide features such as real time alerts for predefined rules, forensic search capabilities, statistical analysis of crowd and traffic flow and more. Ideally the user would like to have some form of VCA functionality for EVERY camera deployed. In an urban environment the VCA system is required to handle a variety of detection tasks dealing with people, vehicles and objects and their behaviors and interactions with each other. Given unlimited computational resources this task is still a very challenging one and is expected to require significant research from the industry and academia in the foreseeable future. In a real life deployment where cost is a major factor the availability of sufficient computing resources for VCA becomes a limiting factor which may severely limit the usability of such a technology in a large scale installation. By nature of being a complementary system component, a VCA system deployment is expected to provide good detection performance (high POD and low FAR) while maintaining the following critical conditions: (a) acceptable cost relative to the other components and ( b) minimal impact on the performance of other system features. Additional considerations for such a deployment are: (c) complexity of setting up and configuring the analytics (d) ease of management and maintenance (e) upgrade path. This presentation will analyze the strengths and weaknesses of known VCA deployment architectures namely based and based in the context of a large scale deployment scenario and will demonstrate an alternative architecture developed and patented by Agent Vi. This proven architecture known as Image Processing Over IP networks (IPoIP) enables providing the end user with a system that scores very highly on all points (a)-(e) mentioned above. The cost/performance advantage of IPoIP is achieved through distribution of the VCA task between the edge device (IP camera or video encoder) and a server. The edge device is tasked with performing the initial analysis of the video stream and extracting information which is relevant in the context of video scene analysis. This information sent to the server as a continuous data stream where it is further analyzed and turned into metadata describing the objects in each and every camera view. The generated metadata is recorded for later offline search functionality and also analyzed in real time to detect deviation from any user defined rule and if this is detected an appropriate event is generated and distributed to any listening client. Using this architecture it is possible to support full VCA functionality for all cameras in a large surveillance installation with a minimal cost overhead while maintaining very high and feature rich performance. BIOGRAPHY