Automatic segmentation of objects of interest in video: a unified framework

Automatically extracting interesting objects from videos is a very challenging task. Traditional video segmentation algorithms assume that objects of interest are the moving objects, whereas they neglect the fact that some stationary objects are also likely to attract human interest. To remove this limitation, this study provides a unified video object segmentation framework by taking human visual attention perception into account. It is implemented by three major steps. First, the JSEG algorithm (Y. Deng et al, IEEE Trans. Pattern Anal. Machine Intell., vol.23, p.800-810, 2001) is adopted to partition each frame into homogenous regions. Following that, based on a number of visual features that have been proven to be able to influence attention, two visual attention models are proposed to calculate the attention value for each region. Finally, a combination of the spatial segmentation and the visual attention models produces all the objects of interest with and without motion. Simulation results on standard video sequences demonstrate its effectiveness.