Three-Dimensional Foot Position Estimation Based on Footprint Shadow Image Processing and Deep Learning for Smart...

Area of Science:

• Biomedical Engineering
• Computer Vision
• Sports Technology

Background:

• The digital fitness market is rapidly growing, with smart home exercise equipment like trampolines gaining traction.
• Accurate user motion recognition is crucial for effective self-guided exercise systems in smart fitness equipment.
• Previous methods for motion tracking faced challenges with calibration and environmental variations.

Purpose of the Study:

• To develop a deep learning-based system for estimating the 3D foot positions using footprint shadow images from a smart trampoline.
• To overcome limitations of prior approaches, such as geometric calibration and sensitivity to illumination changes.

Main Methods:

• A smart trampoline system with an upward-looking wide-angle camera and an embedded processing board was utilized.
• A modified Fast-RCNN network based on ResNet-50 was employed for end-to-end deep learning without calibration.
• The region proposal network was adapted for location regression, distinct from bounding box regression.

Main Results:

• The proposed deep learning algorithm achieved high accuracy in estimating 3D foot positions.
• Root mean square errors for X, Y, and Z directions were 8.32 mm, 15.14 mm, and 4.05 mm, respectively.
• The system demonstrated effectiveness in motion recognition and performance evaluation for jumping exercises.

Conclusions:

• The developed deep learning system accurately estimates 3D foot positions from footprint shadows on a smart trampoline.
• This technology enables enhanced motion recognition and performance evaluation for home-based fitness and rehabilitation exercises.
• The calibration-free, end-to-end learning approach offers a robust solution for smart fitness equipment.