A novel two-level shape descriptor for pedestrian detection

The demand for pedestrian detection and tracking algorithms is rapidly increasing with applications in security systems, human computer interaction and human activity analysis. A pedestrian is a person standing in an upright position. Previous work involves using various types of image descriptors to detect humans. However, the existing approaches, although exhibit low misdetection rate, result in high rate of false alarms in the case of complex image backgrounds. In this work, a novel approach for pedestrian detection is proposed which is based on the combined use of two object detection approaches with the aim of reducing the false alarm rate of the individual detectors. These are the Histogram of Oriented Gradients (HOG) and a Shape Context based object detector (SC). Preliminary results are very encouraging and demonstrate clearly the ability of the proposed system to reduce the number of false alarms without significant increase in the processing time.