Effective pedestrian detection using SVDD-based criterion for region integration

Pedestrian detection is one of the most important techniques for surveillance applications. This paper proposes an effective method for pedestrian detection in low-contrast images. The main characteristic of the proposed method is a two-stage moving object extraction. In the first stage, the watershed algorithm is used to extract multiple regions of moving objects. In the second stage, a novel criterion is introduced to integrate the segmented moving object regions. Specifically, the criterion is calculated on the basis of the distance from a center of the support vector data description (SVDD), where its hypersphere is constructed by using pedestrian features. By monitoring this SVDD-based criterion for the region integration, the segmented regions are appropriately integrated based on pedestrian features. This two-stage approach can extract the moving objects in low-contrast images and improve the performance of the pedestrian detection. Experimental results have demonstrated the effectiveness of the proposed method.