Parametric & non-parametric background subtraction model with object tracking for VENUS

Background subtraction is the process by which we segment moving regions in image sequences. These moving regions refer only to movements in foreground objects. A background segmentation model is one that will not detect movements in the background or of objects that belong to the background. It is common to model the background using a Normal (Gaussian) distribution over its pixel's intensity values however Normal distribution models fail to incorporate the multi-modal intensity distribution. This thesis explores background subtraction techniques with the help of Normal Distributions that use parameters such as the learning rate as well as a segmentation model without any parameters. Both models are compared for accuracy and efficiency of segmentation. Segmented pixels are then grouped into objects and tracking of these objects is performed using the properties of these groups of pixels. These results are then provided as input to the first phase of the modular Video Exploitation and Novelty Understanding System (VENUS) which is used for understanding novelty (or abnormal events) in streams of video.