Implementation of HOG based feature extraction method

Abstract

Human detection on emerging intelligent transportation systems is a challenging task in hardware implementation. The Histogram of Oriented Gradients (HOG)-based human detection is the most successful algorithm due to its superior performance. Unfortunately, more intensive computations and poor performance at a multi-scale and low-contrast make human detection more difficult and unreliable. To address the aforementioned problems, an efficient histogram of edge-oriented gradients-based human detection is proposed to preserve the edge gradients at low-contrast and support multi-scale detection. The proposed algorithm uses approximation methods and adopts a pipelined structure that utilizes low-cost and high-speed, respectively. Experiments conducted on various challenging human datasets show that the proposed method provides efficient detection. This algorithm has been synthesized on Xilinx Spartan 3 FPGA software and board, achieving better hardware utilization compared to other state-of-the-art approaches.