Use of Sparse Representation for Pedestrian Detection in Thermal Images

Pedestrian detection plays a paramount role in advanced driver assistant system (ADAS) and autonomous vehicles, especially with the growth of aging population. The purpose of pedestrian detection is to identify and locate people in a dynamic scene or environment. It needs to tackle the challenges such as illumination, color, texture, clothing, and background complexities. Different from visible imaging system, thermal imaging depends on objects' emissivity, and thus has the advantage on discriminating human body from the cool background. In this study, sparse representation is proposed for pedestrian detection in thermal images. Two types of dictionaries, i.e. a generic dictionary optimized by K-SVD and a naive dictionary with basis atoms being directly composed of training samples, are employed to represent image features. In the implementation, a boundary box shrinking scheme is applied to improve the accuracy of the detection through finding proper size for the boundary box. The experimental results demonstrate a comparable performance of the proposed approach.