Limb movement detection and analysis based on visual recognition of human posture

Zhiguo Xiao, Chunxiang Wang, Tianjiao Ding, Xiangfeng Shen, Xinyuan Li, Dongni LiDiscov. Artif. Intell 2025 · DOI

Abstract

This paper introduces a high-precision, real-time framework for human motion analysis designed to bridge the gap between general pose tracking and personalized medical rehabilitation. By integrating an optimized distance-weighted K-Nearest Neighbors algorithm (DW_KNN*) with a temporal prediction model (CALstm), the system achieves 98.2% accuracy in action recognition. The framework provides tailored exercise guidance by analyzing historical performance, offering a proactive approach to physical therapy.

Introduction

While computer vision has made significant strides in human pose estimation, applying these tools to physical therapy and fitness remains challenging. Current systems often suffer from high computational costs, making real-time feedback difficult on consumer-grade hardware. Furthermore, most systems lack personalization, comparing users against a rigid "perfect" template while ignoring individual physical limitations or long-term recovery trajectories. This research addresses these gaps by creating a lightweight, intelligent system capable of real-time interaction and personalized rehabilitation forecasting.

Methodology

The researchers developed a multi-stage pipeline that transforms raw video input into actionable medical insights through several key components:

  • Pose Extraction: The system utilizes BlazePose to track 33 skeletal key points. It operates at roughly 10x the frame rate of OpenPose while maintaining high accuracy, making it ideal for mobile or home-based applications.
  • Feature Engineering: Instead of raw coordinates, the system uses 22 pairs of feature vectors representing specific limb movements and body bending directions. This vectorization bypasses heavy image-processing layers during classification, drastically increasing speed.
  • DW_KNN* Algorithm: This optimized algorithm improves standard KNN by focusing on the most relevant limb vectors for specific exercises and using weighted Chebyshev distance to prioritize neighbors in the feature space.
  • Action Evaluation: Posture vectors are compared against a standard database using cosine similarity. The system applies Exponential Moving Average (EMA) smoothing to filter out visual noise and jitter from the video feed.
Flow of the evaluation and feedback mechanism
  • CALstm (Temporal Prediction): A hybrid model consisting of a CNN, an LSTM, and an Attention mechanism. This model analyzes longitudinal data—such as the "Motion Time Dataset" containing 8,000 exercise records—to predict a user’s performance for the following day.

Key Results and Discussion

The proposed DW_KNN algorithm achieved a classification accuracy of 98.21%*, significantly outperforming classic KNN (71.85%), Random Forest, and Naive Bayes on the study's Rehab dataset.

Real-time action recognition and counting interface

In terms of predictive precision, the CALstm model achieved a Mean Absolute Error (MAE) of 1.383 for motion count and 0.508 for motion time. This indicates the model is remarkably accurate at forecasting how many repetitions a patient can perform and for how long. By differentiating weights for upper and lower limb movements, the system effectively reduces the "distance value space" the model must navigate, further enhancing accuracy.

Conclusion and Future Work

While the system provides robust real-time feedback, it faces challenges with depth noise in standard RGB camera data, particularly for movements occurring along the camera's Z-axis. Future research will focus on improving performance in varied environmental conditions and integrating the framework with VR and AR devices to create more immersive, gamified rehabilitation scenarios.

Summary updated

BibTeX
@article{xiao2025limb,
  author = {Zhiguo Xiao and Chunxiang Wang and Tianjiao Ding and Xiangfeng Shen and Xinyuan Li and Dongni Li},
  title = {{Limb movement detection and analysis based on visual recognition of human posture}},
  journal = {Discov. Artif. Intell.},
  year = {2025},
  doi = {10.1007/s44163-025-00246-4},
}
Figure