Banner occlusion detection in security surveillance video

In security surveillance video (SSV), foreign object occlusion is increasingly common. Automatic detection of suspicious occlusion has become important. In this paper, a banner occlusion detection approach is proposed. The proposed approach first detects the banner in the image using both color feature and shape feature. More specifically, the proposed approach exploits the HSV color space to extract the suspected banner area, which is further refined using mathematical morphology, corner detection and boundary points sampling based line-fitting. Then, the proposed approach exploits area ratio and banner location to determine banner occlusion when a banner exists. Finally, a Gradient Boost Decision Tree is constructed to learn features and predict whether there exists occlusion. Experiments are conducted using real surveillance video to validate the effectiveness of the proposed algorithm.