Foreground Background Segmentation using Temporal and Spatial Markov Processes

Markov Processes Pankaj Kumar and ¤Kuntal Sengupta Department of Electrical and Computer Engineering, National University of Singapore 10 Kent Ridge Crescent, Singapore 119260 e-mail: [email protected] EDICS: 2-SEGM, 2-SEQP November 6, 2000 Abstract In this paper, we present a novel approach in extracting interesting foreground regions from a complex and largely stationary background scene. The method is a substantial extension to the existing background subtraction techniques. The advantage of the proposed technique are the following: it removes false detection due to shadow, due to illumination changes caused by Automatic Exposure Correction (AEC), and ensures temporal and spatial contiguity in the results. A background pixel is ̄rst statistically modelled by two Gaussian distributions. These statistical estimates are used next as important parameters in test for spatial and temporal contiguity in the foreground regions. Temporal contiguity of pixels are ensured by modelling the pixel value as a Markov Random Sequence and the Bayes smoothing is applied on the result of the foreground/background hypothesis test. Spatial contiguity is ensured by using Markov Random Fields. We make an analogy of a pixel process in a neighborhood of pixels and a lattice site in a lattice structure. The spatial constraints are applied by applying the Gibb's energy distribution function. The segmentation results are quite fast (about 3 frames a second on a desktop PC ¤Corresponding author