Abstract
We consider cameras whose outputs do not reflect true scenes as faulty cameras. To build a fault detection video surveillance system without using additional hardware devices, we use the video outputs from cameras to do self-checking. We study two categories of faults: Spatial Faults and Temporal Faults, and reduce the two sub-problems into graph theoretical problems on two graphs (Surveillance Sharing Graph (SSG) and Surveillance Partitioning Graph (SPG)). We prove a theoretical upper bound for the spatial fault detection, and develop two algorithms for detecting the two types of faults respectively.Then, we integrate both in a framework, which is capable of the following: Given the outputs from a video surveillance system, it isolates cameras which are faulty or suspected to be faulty. It gives warnings with types of faults, locations, detection confidence. Our experiments confirm the effectiveness of the framework's methodologies.