Steering kernel-based video moving objects detection with local background texture dictionaries

Texture and color information as important characters of video objects have been widely used in the detection of moving objects. A detection algorithm based on texture may cause detection errors in regions of blank texture and heterogeneous texture, and a detection algorithm based on color is easily influenced by illumination changes and shadows. In this paper, a new detection and fusion algorithm is proposed. At the detection stage based on texture, the background texture is classified according to the steering kernel. At the fusion stage, for the moving objects detected on the basis of texture and color respectively, a scheme based on a boundary selection strategy is proposed for combining the different detection objects. A relatively smooth boundary is selected as the true boundary, and the shadow detection is carried out to assist the boundary selection. Experimental results verify the advantages of the proposed algorithm as compared to the existing state-of-the-art algorithms.

Figure optionsDownload as PowerPoint slideHighlights
► The steering kernel is used to describe the background texture robustly.
► With the steering kernel, different background texture dictionaries are constructed to represent background texture.
► The smoother boundary is selected as true boundary.
► After the shadow detection, the weighting curvature sum is defined to determine the true boundary in the shadow region.