HRDepthNet: Depth Image-Based Marker-Less Tracking of Body Joints

Area of Science:

• Computer Vision
• Machine Learning
• Biomedical Engineering

Background:

• Existing human joint detection methods primarily use color images, neglecting depth images' advantages like light-invariance and texture independence.
• Depth images offer unique benefits for human pose estimation, particularly in uncontrolled lighting or texture-poor environments.

Purpose of the Study:

• To introduce High-Resolution Depth Net (HRDepthNet), a deep learning model for human joint detection using solely depth images.
• To adapt and retrain the established HRNet architecture for effective processing of depth data.
• To create and utilize a specialized dataset for training and evaluating HRDepthNet on depth image-based human pose estimation.

Main Methods:

• Retraining the HRNet architecture using a custom dataset of depth images from the Timed Up and Go test.
• Manual annotation of RGB images corresponding to the depth data for ground truth.
• Evaluation using COCO metrics for joint detection accuracy and analysis of positional errors.

Main Results:

• HRDepthNet demonstrated superior performance in human joint detection compared to HRNet applied to corresponding RGB images.
• The depth-based model achieved high accuracy, with median positional deviations of 1.619 cm (x-axis), 2.342 cm (y-axis), and 2.4 cm (z-axis).

Conclusions:

• HRDepthNet provides a robust and accurate solution for human joint detection in depth images.
• This approach offers significant advantages over color image-based methods, especially in challenging environmental conditions.
• The developed model shows promise for applications in healthcare, particularly in geriatric assessments and motion analysis.