Optimising resource allocation for background modeling using algorithm switching

Background maintenance is a complex problem due to varying scene conditions. Typically, a single algorithm cannot handle the complex scene changes that occur in visual surveillance applications. Also complex background modelling techniques, for example mixture of Gaussians have a high computational and communication demand compared to simple techniques such as a uni-model background model or simple frame-differencing with an adaptive threshold. In this paper we present an algorithm switching approach that can handle multi-model background scenes. Starting with a less computational and less data intensive uni-model background subtraction algorithm the system switches to a complex multi-model background subtraction there by saving valuable software and hardware resources both in terms of power and computation time. Our results show this algorithm switching approach can be used to effectively handle various scene conditions encountered in real time surveillance systems with optimal use of system resources.