A Lightweight Pedestrian Detection Engine with Two-Stage Low-Complexity Detection Network and Adaptive Region Focusing Technique

Pedestrian detection has been widely used in applications such as video surveillance and intelligent robots. Recently, deep learning-based pedestrian detection engines have attracted lots of attention. However, the computational complexity of these engines is high, which makes them unsuitable for hardware- and power-constrained mobile applications, such as drones for surveillance. In this paper, we propose a lightweight pedestrian detection engine with a two-stage low-complexity detection network and adaptive region focusing technique, to reduce the computational complexity in pedestrian detection, while maintaining sufficient detection accuracy. The proposed pedestrian detection engine has significantly reduced the number of parameters (0.73 M) and operations (1.04 B), while achieving a comparable precision (85.18%) and miss rate (25.16%) to many existing designs. Moreover, the proposed engine, together with YOLOv3 and YOLOv3-Tiny, has been implemented on a Xilinx FPGA Zynq7020 for comparison. It is able to achieve 16.3 Fps while consuming 0.59 W, which outperforms the results of YOLOv3 (5.3 Fps, 2.43 W) and YOLOv3-Tiny (12.8 Fps, 0.95 W).