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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Embedded Systems

Background:

• Deep learning models for pedestrian detection offer high accuracy but suffer from significant computational complexity.
• This complexity limits their deployment on hardware- and power-constrained mobile platforms like surveillance drones.
• Existing methods often fail to balance detection performance with computational efficiency for edge devices.

Purpose of the Study:

• To develop a lightweight pedestrian detection engine suitable for mobile and embedded applications.
• To reduce computational complexity and power consumption without sacrificing detection accuracy.
• To enable efficient pedestrian detection on resource-limited platforms such as drones.

Main Methods:

• Proposed a two-stage low-complexity detection network combined with an adaptive region focusing technique.
• Developed a lightweight pedestrian detection engine with significantly reduced parameters (0.73 M) and operations (1.04 B).
• Implemented and compared the proposed engine with YOLOv3 and YOLOv3-Tiny on a Xilinx FPGA Zynq7020.

Main Results:

• Achieved comparable precision (85.18%) and miss rate (25.16%) to existing pedestrian detection designs.
• The proposed engine demonstrated superior performance on the FPGA, reaching 16.3 Frames Per Second (Fps) at 0.59 Watts (W).
• Outperformed YOLOv3 (5.3 Fps, 2.43 W) and YOLOv3-Tiny (12.8 Fps, 0.95 W) in terms of speed and power efficiency.

Conclusions:

• The proposed lightweight pedestrian detection engine effectively addresses the computational challenges of deep learning models on mobile devices.
• It offers a viable solution for real-time pedestrian detection in power- and hardware-constrained applications like drone surveillance.
• The engine provides a significant advancement in balancing performance and efficiency for edge AI applications.