Video-Based Lifting Action Recognition Using Rank-Altered Kinematic Feature Pairs.

Area of Science:

• Biomechanics
• Computer Vision
• Ergonomics

Background:

• Traditional lifting action recognition relies on wearable sensors and deep learning, which can be resource-intensive.
• Implementing these methods on limited hardware poses significant challenges for real-time monitoring.

Purpose of the Study:

• To develop a streamlined, computer vision-based method for identifying and counting lifting actions in videos.
• To enable reliable real-time monitoring of lifting tasks for injury prevention.

Main Methods:

• Utilized BlazePose for real-time human keypoint detection.
• Extracted and preprocessed kinematic features from detected keypoints.
• Employed rank-altered kinematic feature pairs for lifting/non-lifting classification.
• Developed a lifting counting algorithm based on the classification predictions.

Main Results:

• An ensemble classifier using nine key rank-altered kinematic feature pairs achieved over 0.89 in accuracy, precision, recall, and F1 score.
• The lifting counting accuracy reached 0.90 with a minimal latency of 0.06 ms.
• The proposed method demonstrated at least 12.5 times faster performance compared to baseline classifiers.

Conclusions:

• Computer vision-based kinematic features offer an effective and efficient solution for recognizing lifting actions.
• The developed method is suitable for deployment on resource-constrained platforms like mobile and embedded systems.
• Real-time monitoring of lifting tasks can proactively prevent work-related low-back injuries.