Abstract
Human Motion Analysis (HMA) is critical for applications ranging from sports medicine to industrial ergonomics. This paper introduces a low-cost, modular wearable system employing Inertial Measurement Units (IMUs) to classify complex human motions. By implementing a robust five-stage data conditioning pipeline—including Kalman filtering and novel distance-based feature extraction—the system achieves high classification accuracy. Evaluation across hand gestures and diverse walking patterns shows that while traditional machine learning models are effective, Deep Learning architectures like Long Short-Term Memory (LSTM) networks can reach 100% accuracy, providing a portable and robust alternative to fixed vision-based motion capture.
Introduction
Traditional Human Motion Analysis (HMA) primarily relies on vision-based systems. While accurate, these systems are expensive, restricted to controlled indoor environments, and prone to "blind spots" where the line of sight is obstructed. Existing wearable alternatives often suffer from being bulky or limited to specific limbs.
The researchers address these limitations by developing a system that is modular, robust, and preventative. The goal is to monitor repetitive tasks in "natural scenes"—such as factories or outdoor slopes—to identify risks for Work-Related Musculoskeletal Disorders (MSD) among workers like nurses or factory staff.
Methodology
The system architecture consists of a hardware-software ecosystem designed for flexibility. The hardware utilizes Arduino Nano boards interfaced with MPU 6050 IMUs (accelerometers and gyroscopes) and SD card modules for data logging.
The core of the system is a five-stage data conditioning pipeline:
- Signal Power Filtering: Filters out sensors with insignificant signal fluctuations.
- Clustering Index Filtering: Selects sensors providing optimal separation between movements.
- Kalman Filtering: Uses an Extended Kalman Filter to fuse sensor data, reducing noise and estimating limb orientation via quaternions.
- Distance-Measure Construction: A novel step calculating Euclidean or Dynamic Time Warping (DTW) distances between live and reference signals to enrich the feature set.
- Principal Component Analysis (PCA): Compresses high-dimensional data while preserving critical information.
The system was validated using two datasets: a Hand Gesture set (13 movements) and a Walking Pattern set (15 scenarios). The walking patterns included three terrains (plain, slope, stairs) and three shoe types (sport shoes, safety boots, and high heels) to test the system's sensitivity to external attire and environment.
Results and Insights
The modularity of the system allows for targeted monitoring of specific limbs or holistic full-body tracking, as shown in the deployment views.
Key findings include:
- Classification Accuracy: For hand gestures, traditional Logistic Regression achieved 92.14% accuracy. For walking patterns, Naive Bayes reached 76.24%. However, the LSTM neural network outperformed all traditional models, achieving 100% accuracy in both categories.
- Feature Engineering: The "distance measures" introduced in Stage IV significantly boosted performance by transforming temporal signal data into a format more easily processed by ML algorithms.
- Environmental Robustness: The system effectively distinguished between subtle differences in gait caused by different terrains and footwear, proving its utility for real-world injury risk assessment.
- Data Efficiency: The pipeline reduced data dimensionality by over 7-fold while improving the identification of subtle limb movements.
Conclusion and Future Work
While the system demonstrates exceptional accuracy, several areas for improvement remain. Currently, data is logged to SD cards, which introduces latency. Future iterations will focus on wireless streaming (Bluetooth/Wi-Fi) for real-time biofeedback. Additionally, future research will involve more diverse subject groups and detailed power consumption analysis to ensure the device's viability for prolonged industrial monitoring. The ultimate goal is to visualize granular gait parameters, such as joint angles and velocity, to provide deeper medical insights.
Summary updated