An improved pedestrian detection approach for cluttered background in nighttime

Abstract

Pedestrian detection is one of the most interesting topics in driver assistant systems. In a normal two-step detection framework: image segmentation (thresholding) and recognition, the pedestrian areas usually connect with other objects after segmentation, especially in cluttered nighttime images. The bad segmentation result causes the recognition module not to identify the pedestrians. This paper presents a fast template matching approach to locate the most pedestrian-like areas (candidates) in the complex background. At most of the time, the template matching method produces too many non-human candidates. However, our approach employs a set of efficient and simple filters to reject most of unwished candidates to reduce false alarm rate. Experiments show that the proposed method can segment the pedestrian areas well and promote the ability of the pedestrian detection system.