Using deep neural networks for kinematic analysis: Challenges and opportunities.

Area of Science:

• Biomechanics
• Computer Vision
• Artificial Intelligence

Background:

• Traditional kinematic analysis relies on lab-based optical cameras and reflective markers.
• Advancements in AI, particularly deep neural networks, enable markerless motion analysis.
• Markerless techniques expand the feasibility of kinematic analysis to outdoor environments.

Purpose of the Study:

• To summarize 2D markerless approaches for estimating joint angles.
• To highlight the strengths and limitations of current markerless pose estimation methods.
• To discuss the potential applications and future improvements in markerless kinematic analysis.

Main Methods:

• Utilizes "pose estimation" algorithms based on supervised learning with neural networks.
• Employs transfer learning to adapt pre-trained models, reducing the need for extensive manual data annotation.
• Focuses on detecting anatomical landmarks for joint angle estimation.

Main Results:

• Markerless systems, despite ongoing development, demonstrate sufficient accuracy for certain applications like coaching and rehabilitation.
• Manual annotation for supervised learning presents limitations including subjectivity, required expertise, and data quality/quantity issues.
• Transfer learning significantly reduces the number of training images required for neural network models.

Conclusions:

• Markerless 2D kinematic analysis is becoming increasingly viable, offering a cost-effective alternative to traditional methods.
• Future improvements may involve novel approaches and physiological constraints to enhance accuracy.
• The technology holds potential for widespread deployment both in laboratory and real-world settings.