Coding-based hough transform for pedestrian detection

In this paper, we present a coding-based Hough transform framework for pedestrian detection in natural images. In traditional Hough transform methods for pedestrian detection, the voting element is represented by a linear combination of codebook entries with uniform probability, which often leads to uncertainty of the reconstruction error in the Hough voting process. To minimize the reconstruction error, we construct a new Hough transform framework to encode the voting element by using the locality-constrained linear coding (LLC) algorithm. Consistently, the voting element casts weighted votes into the Hough image according to its coding coefficients. We evaluated the proposed method on two publically available datasets, namely, the INRIA pedestrian, and TUD Brussels datasets. Experimental results demonstrated the effectiveness of the proposed method.