Abstract
Foreground segmentation is a critical early step in most human-computer interaction applications notably in action and gesture recognition domain. In this paper, an approach to model background which based on luminance-invariant color with an adaptive Gaussian mixture is proposed to discriminate foreground object from their background in complex scene. Firstly, the background model is learned based on the spectral properties of shadows and scene activity. Secondly, the shadow with the hypotheses on color invariance is adaptively set up and updated. Finally, the log-likelihood measurement is to conduct the adaptation. Our experiments are performed on a wide range of practical applications of gesture and action recognition videos. Additionally, the proposed approach is efficient and more robust than premature state-of-the-art with no sacrificing real-time performance.