Histogram of Silhouette Direction code: An efficient HOG-based descriptor for accurate human detection

Histograms of Oriented Gradients (HOG) is an effective means for the human detection from image or video. Based on the HOG principle, this paper presents a more efficient method named Histogram of Silhouette Direction (HSD). To extract the silhouette pattern of human body, the average background model is used. A novel function that combined with brightness and color information is proposed to segment the foreground accurately and adaptively. The histogram of direction code is constructed via Freeman Chain Code of Eight Directions (FCCE) along the extracted silhouette and used as the feature descriptor. Compared with traditional HOG descriptor which calculates gradients of all image pixels, the proposed HSD descriptor can reduce the feature dimension and whole computation greatly. Experimental results with real video sequences show its improvements in both segmentation efficiency and accuracy.