KDE-Based Simultaneous Background Model Learning and Entropy-Based Fusion of Cascaded Features for Video Object Segmentation With Shadow Removal

Object detection with shadow removal is one of the challenging issues in computer vision. Dynamic shadow resembles a moving object’s properties, so separating this shadow from the object is a challenging task. This dynamic shadow if not eliminated, distorts the shape of the object. In this paper, a novel scheme for moving object detection and shadow removal is proposed based on the background modeling in fused feature space, and these models learn to take care of the scene dynamics. Initially, in KDE space, temporal modeling of the spatial KDE (TMS-KDE) is carried out and cascaded features of Gabor and HOG are obtained. Besides, the original video frame is transformed into YCbCr color space and LBP features are extracted. The LBP and cascaded features are fused probabilistically to generate fused feature frames which are used in background modeling. The weights for the feature fusion are determined by the proposed entropy based measure. Background modeling and model learning is a pixel based approach and the pixel is classified as either background or foreground during the learning process. We have tested our proposed method on a wide range of datasets which includes ATON-CVRR, LASIESTA, CD-net, Kaggle, PETS 2006, SGM-RGBD, SBMI 2015, SBMnet 2016 and VIRAT. The proposed scheme is found to take care of different shadow conditions while detecting the moving object. The performance of the proposed scheme is found to be superior to that of many existing schemes.