An adaptive background modeling for foreground detection using spatio-temporal features

Background modeling is a well accepted foreground detection technique for many visual surveillance applications like remote sensing, medical imaging, traffic monitoring, crime detection, machine/robot vision etc. Regardless of simplicity of foreground detection concept, no conventional algorithms till date seem to be able to concurrently address the key challenges like illumination variation, dynamic background, low contrast and noisy sequences. To mitigate this issue, this paper proposes an improved scheme for foreground detection particularly addresses all the aforementioned key challenges. The suggested scheme operates as follows: First, a spatio-temporal local binary pattern (STLBP) technique is employed to extract both spatial texture feature and temporal motion feature from a video frame. The present scheme modifies the change detection rule of traditional STLBP method to make the features robust under challenging situations. The improvisation in change description rule reflects that to extract STLBP features, the mean of the surrounding pixels is chosen instead of a fixed center pixel across a local region. Further, in many foreground detection algorithms a constant learning rate and constant threshold value is considered during background modeling which in turn fails to detect a proper foreground under multimodal background conditions. So to address this problem, an adaptive formulation in background modeling is proposed to compute the learning rate (alpha(b)) and threshold value (T-p) to detect the foreground accurately without any false labeling of pixels under challenging environments. The performance of the proposed scheme is evaluated through extensive simulations using different challenging video sequences and compared with that of the benchmark schemes. The experimental results demonstrate that the proposed scheme outperforms significant improvements in terms of both qualitative as well as quantitative measures than that of the benchmark schemes.