Feature Fusion Based Background Model Learning for Video Object Detection

In this paper, a novel feature fusion based strategy is proposed to model complex background to effectively detect the foreground in a real world scene. In order to take care of bad weather conditions of a scene having low visibility and dynamic entities in the background, we have proposed "m-smoothing" filter to obtain the local features. This local feature is fused with the Local Binary Pattern (LBP) based feature to generate a new feature frame for background modeling. Background model learning takes place in the fused feature space to extract the foreground. The proposed method is tested with change detection data sets (Blizzard and Wet Snow) and the performance of our proposed method outperform LBP, STLBP and C-EFIC based methods.