Real-Time Energy Efficient Hand Pose Estimation: A Case Study.

Area of Science:

• Computer Vision
• Machine Learning
• Hardware Acceleration

Background:

• Human hand pose estimation is crucial for applications like virtual reality and in-air signature verification.
• Current state-of-the-art methods use Convolutional Neural Networks (CNNs) on Graphics Processing Units (GPUs), which are power-intensive and not portable.
• There is a need for energy-efficient and portable solutions for real-time hand pose tracking.

Purpose of the Study:

• To develop an energy-efficient solution for depth camera-based 3D hand pose estimation.
• To compress deep neural network models without sacrificing accuracy.
• To design and implement a custom hardware architecture for portable hand pose estimation.

Main Methods:

• Applied dynamic quantization techniques to compress the deep neural network model.
• Designed a custom hardware architecture targeting Field-Programmable Gate Arrays (FPGAs) for high energy efficiency.
• Implemented the compressed model on a Xilinx UltraScale+ MPSoC FPGA.

Main Results:

• Achieved significant model compression through dynamic quantization.
• The FPGA implementation demonstrated a 4.2x speed increase compared to the original GPU implementation.
• The solution achieved a 577.3x improvement in energy efficiency over the NVIDIA GeForce GTX 1070.

Conclusions:

• The proposed method offers a substantial improvement in speed and energy efficiency for 3D hand pose estimation.
• FPGA-based acceleration is a viable approach for creating portable and power-efficient hand tracking systems.
• This work enables practical deployment of advanced hand pose estimation in mobile and embedded devices.