Lightweight and Energy-Efficient Deep Learning Accelerator for Real-Time Object Detection on Edge Devices.

Area of Science:

• Embedded Systems
• Machine Learning
• Hardware Acceleration

Background:

• The Internet of Things (IoT) is rapidly expanding, driving the need for Tiny Machine Learning (TinyML) solutions.
• Existing deep learning algorithms are often too complex and power-hungry for resource-constrained IoT devices.
• Current hardware accelerators are not suitable for embedded, real-time processing on edge devices.

Purpose of the Study:

• To develop a compact, hardware-optimized deep learning model for real-time inference on IoT devices.
• To propose a lightweight and energy-efficient hardware architecture for deep learning acceleration.
• To enable efficient object detection on battery-operated, resource-constrained edge devices.

Main Methods:

• Model simplification and compression techniques were applied to create an optimized network model.
• A novel hardware architecture was designed for a lightweight and energy-efficient deep learning accelerator.
• Experiments were conducted on a Xilinx ZC702 FPGA operating at 100 MHz.

Main Results:

• The optimized model successfully performed object detection.
• The proposed hardware design achieved 1.25x smaller logic and 4.27x smaller BRAM size compared to previous works.
• Energy consumption was reduced by approximately 10.37x, with real-time processing at 43.95 fps.

Conclusions:

• The presented optimized model and hardware architecture are suitable for real-time TinyML applications on IoT devices.
• The solution significantly reduces hardware footprint and energy consumption, addressing key limitations of current edge computing.
• This work facilitates the deployment of advanced AI capabilities on low-power, embedded systems.