Efficient Body Motion Quantification and Similarity Evaluation Using 3-D Joints Skeleton Coordinates

Aouaidjia Kamel, Bin Sheng, Ping Li, Jinman Kim, David Dagan FengIEEE Trans. SMC: Systems 2021 · DOI

Abstract

Analyzing human motion is complex due to the articulated nature of the skeleton and variations in individual performance. This paper introduces a robust framework for quantifying and comparing human body movements using raw 3D skeleton coordinates. The proposed method decomposes motion into "global" and "local" components and utilizes the Kabsch algorithm for transformation estimation. This approach provides a movement similarity evaluation that is independent of a person's height, distance from the camera, or specific action labels. Experimental results on public and custom datasets demonstrate high discriminative power and accurate temporal alignment.

Introduction

The analysis of human motion is a significant challenge in computer vision and biomechanics. Because the human skeleton is articulated, a movement in one joint (e.g., the shoulder) causes displacement in distal joints (e.g., the wrist), even if those distal joints remain stationary relative to their parent. This ambiguity makes it difficult to determine whether a joint's movement is due to its own action or the movement of a parent joint.

While existing systems often focus on action recognition—classifying movements into categories like "walking" or "jumping"—they frequently lack the precision required for quantification. Measuring how a movement is performed is critical for applications such as physical therapy, sports coaching, and high-fidelity human-computer interaction. This paper addresses the need for precise metrics to compare movement similarity across different subjects and environments.

Methodology

The proposed framework consists of two primary stages: Motion Quantification and Motion Comparison.

Figure 4: The overall framework for motion quantification and similarity evaluation, illustrating the pipeline from raw skeleton sequences to the final similarity percentage.

Hierarchical Motion Decomposition

To resolve the ambiguity of articulated movement, motion is categorized into two types:

  1. Global Motion: The movement of a joint relative to a fixed "root" point, typically the hip center (j1).
  2. Local Motion: The movement of a joint relative to its immediate parent in the skeletal hierarchy (e.g., the wrist relative to the elbow). This isolates the specific contribution of each individual joint.
Figure 2: (a) The 15-joint skeletal model used for tracking. (b) Decomposition of motion into local and global components, showing rotation and translation metrics.

Transformation Estimation and Limb Angles

The system employs the Kabsch algorithm to estimate rotation and translation matrices between consecutive frames. This allows the framework to interpret the direction and orientation of movement directly from raw point data. Additionally, the Law of Cosines is used to calculate eight fundamental angles between limbs (e.g., knee or elbow angles) to capture the posture of the motion.

Comparison Algorithm

A multi-level algorithm compares these metrics across two skeleton sequences. By utilizing adjustable thresholds for rotation, translation, and angles, the system accounts for minor performance variations. The final output is a similarity percentage calculated for each joint, each frame, and the overall motion sequence.

Experimental Evaluation

The method was evaluated using two datasets:

  • UTD-MHAD: A public multimodal dataset containing 27 actions performed by 8 subjects.
  • Custom Kinect V2 Dataset: A dataset recorded for a user study involving four sports movements (basketball shooting, football kick, baseball swing, and tennis racquet swing) performed by five users at varying distances and speeds.

Key Results and Insights

  • Invariance to Setup: A major finding is that the quantification is largely invariant to a user's height and distance from the camera. By normalizing coordinates and focusing on relative transformations, similarity scores remain consistent across different physical setups.
  • Discriminative Power: On the UTD-MHAD dataset, similar actions performed by different individuals typically yielded similarity scores above 70%, while distinct actions (e.g., "walking" vs. "sitting") scored below 40%.
  • Temporal Alignment: Compared to standard techniques like Subsequence Dynamic Time Warping (SS-DTW), the proposed method achieved lower Euclidean distance values, indicating more accurate temporal alignment.
  • Impact of Speed: While height and distance have minimal impact, movement speed significantly affects similarity scores. Because the algorithm compares frames of the same rank, tempo differences lead to lower scores even if the physical paths are identical.

Conclusion

This framework provides an efficient and robust means of evaluating human motion similarity. By isolating local joint movements and utilizing coordinate-independent transformations, it overcomes common hurdles in skeletal motion analysis. Future work will focus on refining threshold sensitivity and implementing advanced temporal warping to better handle significant variations in movement timing. This technology holds promise for real-time feedback in sports training and physical rehabilitation.

Summary updated

BibTeX
@article{kamel2021efficient,
  author = {Aouaidjia Kamel and Bin Sheng and Ping Li and Jinman Kim and David Dagan Feng},
  title = {{Efficient Body Motion Quantification and Similarity Evaluation Using 3-D Joints Skeleton Coordinates}},
  journal = {IEEE Trans. Syst. Man Cybern. Syst.},
  year = {2021},
  doi = {10.1109/tsmc.2019.2916896},
}
Figure