Spatial multimodal mean background model for real-time MTI

Abstract

One of the important tasks in video surveillance is to detect and track targets moving independently in a scene. Most real-time research to date has focused on scenarios from stationary cameras where there is limited movement in the background, such as videos taken at traffic lights or from buildings where there is no background proximal to the background. A more robust method is needed when there are moving background objects such as trees or flags close in the camera or when the camera is moving. In this paper we first introduce a variant of the multimodal mean (MM) background model that we call the spatial multimodal mean (SMM) background model that is better suited for these scenarios while improving the speed of the mixture of Gaussians (MoG) background model. It approximates the multimodal MoG background with the generalization that each pixel has a random spatial distribution. The SMM background model is well suited for real-time nonstationary scenes since it models each pixel with a spatial distribution and the simplifications make it computationally feasible to apply image transformations. We then describe how this can be integrated into a real-time MTI system that does not require the estimation of depth.