Training-Free Skeleton-Semantic Keyframe Extraction for Fixed-View Industrial Assembly Video: Method Design and Case...

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Industrial assembly videos exhibit significant temporal redundancy.
• Existing preprocessing methods often require task-specific labels or retraining, limiting their applicability.
• There is a need for training-free, interpretable methods for video data reduction.

Purpose of the Study:

• To design a deterministic, training-free skeleton-semantic keyframe extraction framework for fixed-view monocular assembly video preprocessing.
• To develop a method that reduces temporal redundancy without requiring task-specific labels or retraining.
• To evaluate the proposed method's effectiveness in terms of semantic-distance variation and pose-state separation.

Main Methods:

• Utilized OpenPose BODY-25 keypoints, applying confidence filtering, normalization, and interpolation to represent upper-body joints.
• Implemented a neck-relative bilateral-wrist activity control for an inverse threshold, alongside sequential reference updates and a maximum-gap rule for output budget control.
• Evaluated performance using metrics like CVsd and MSD, and compared against MViT-V2-S feature clustering and human annotations for transition events.

Main Results:

• The proposed method achieved a CVsd of 0.3340 ± 0.1117 and MSD of 0.7781 ± 0.1003.
• Significantly reduced semantic-distance variation by 68.7% and increased adjacent pose-state separation by 162.8% compared to MViT.
• Retained 79.41% of transition events within ±0.5 s, demonstrating a trade-off between pose-state diversity and process-boundary coverage.

Conclusions:

• The developed framework is an interpretable and budget-controllable preprocessing strategy for pose-oriented review and data reduction in fixed-view assembly videos.
• The skeleton representation is identified as the primary factor for improved organization and separation.
• Further validation in diverse production environments and downstream systems is recommended for broader deployment.