Shadow removing in a surveillance system by a multi-resolution classification strategy

Shadow removing is a key issue for moving objects detection in a surveillance systems. However, few research address this problem by a learning strategy. In this paper, we present a multi-resolution classification method to remove shadows from the object detection result. Because the number of samples which denote the shadow and object is reasonably large, we adopt a coarse-to fine strategy during the classification process. By partitioning feature space into hypercubes according to different resolutions, we train a group of classifiers which can label the samples for testing from coarse to fine. Support Vector Machines are chosen in the process of training and the hypercubes which represent support vectors are subdivided in order to generate the sample set intended for training in a higher resolution. Because of the conglomeration property of the samples to be tested, we can label most of the samples using the simple classifiers trained at low resolution. In some cases, the method presented in this paper can reduce the computational complex of the classification algorithm. Finally, experimental results have substantiated the effectiveness of the proposed method.