Detection and Decomposition of Foreground Target from Image Sequence
Kwok Leung Chan · Machine Vision Applications · 2013
A surveillance video contains two components – the background scene and the foreground targets. Foreground detection is difficult when there are illumination changes and background motions in the scene. We propose a two-step background subtraction framework for foreground detection. The background is modeled as block-based color Gaussian mixture model. In the first step of background subtraction, the current image frame is compared with the background model via a spatial similarity measure. The potential targets are separated from most of the background pixels. In the second step, if a potential target is sufficiently large, the enclosing block is compared with the background model again in order to obtain a refined shape of the foreground. Complex target shape may exhibit multiple motions. Decomposition of the foreground region into meaningful parts is essential for the recognition of the activity. We adopt the morphological shape decomposition algorithm to decompose each foreground region. The method is enhanced by considering color cue in the decomposition process. We test the foreground detection and decomposition methods on a swimming image sequence.