Accurate Human Silhouette Extraction in Video Data by Shadow Evaluation

This paper presents a new shadow detecting method for silhouette extraction of a person in gray-level video sequences. We use a shadow evaluator to verify each raw shadow pixel which was detected by Gaussian distribution analysis. The evaluator considers a raw shadow pixel initially to be a fake shadow pixel, and marks it as a silhouette pixel if it is enclosed or semienclosed by moving occlusion boundaries of a person. Those were extracted by subtracting edges in the current frame from edges of the background. We also propose a silhouette-compensation technique to recover some missing (i.e. removed) silhouette pixels by using a similarity criterion between silhouette pixels and their neighbors. Experimental results show us that the proposed algorithm keeps a silhouette of a person more accurate compared to other methods. Methods advocated by other researchers in YUV or RGB color space, typically remove silhouette pixels as shadow if the color of these pixels is similar to that of the surrounding background. Keywords—Shadow removal, silhouette detection, shadow evaluator, silhouette compensation