Introduction
Physical inactivity remains a global health crisis, yet professional fitness coaching is often inaccessible due to high costs, socioeconomic barriers, and privacy concerns. While technological solutions like computer vision and wearables have attempted to bridge this gap, they face significant hurdles. Computer vision systems suffer from occlusion and privacy issues in public spaces, while wearables are prone to sensor drift and user "logging fatigue." Medical-grade force plates, though accurate, are too bulky and expensive for standard gym environments.
This research introduces LifeMat, a portable, low-cost pressure-sensing mat designed to be "invisible" to the user. By analyzing Center of Pressure (COP) signatures, LifeMat provides a non-invasive, privacy-preserving method for tracking and guiding fitness training with laboratory-grade accuracy.
System Design and Methods
The LifeMat is a 3mm-thick flexible mat embedded with a matrix of 924 piezoresistive sensors. To ensure real-time performance on edge devices, the researchers avoided high-dimensional raw pressure maps in favor of 13 specific Center of Pressure (COP) features. These features are categorized into:
- Displacement: Medial-Lateral and Anterior-Posterior changes in COP position.
- Velocity: The speed of COP movement.
- Acceleration: The rate of change in COP velocity.
- Balance: Foot distance and weight distribution percentage.
The system utilizes "edge-friendly" machine learning classifiers, including Support Vector Machines (SVM) and Random Forest (RF), to identify exercises based on these biometric signatures.
Experimental Setup and Datasets
The study employed two primary datasets:
- Validation Dataset: Data from 16 participants used to benchmark LifeMat against medical-grade force plates.
- Activity Dataset: Collected at a major stadium gym in Tokyo, comprising 8,385 recordings from 30 participants performing four foundational exercises: Squats, Deadlifts, Bicep Curls, and Shoulder Presses.
Results and Discussion
The LifeMat system achieved a peak recognition accuracy of 98.5% using an SVM model. A key finding was that even upper-body exercises like bicep curls produce distinct COP signatures as the body shifts weight to maintain stability.
Analysis of feature importance revealed that Displacement and Balance were the most critical metrics, with displacement alone yielding 98.3% accuracy. Interestingly, adding velocity and acceleration data occasionally decreased performance due to noise in dynamic gym settings. The data also highlighted a "Deadlift Divergence," where heavy bar-based movements required rapid, fine-tuned balance adjustments captured uniquely by the mat's sensors.
Beyond technical performance, qualitative feedback indicated that users, particularly the elderly, appreciated the non-judgmental nature of AI-driven coaching, which reduced the "social complexity" often felt when working with human trainers.
Conclusion and Future Work
LifeMat demonstrates that high-accuracy activity recognition is possible through a low-cost, non-invasive floor interface. Future research will focus on expanding the exercise repertoire to include floor-based movements like planks and transitioning from simple activity recognition to real-time form correction (e.g., detecting back arching during deadlifts). The ultimate goal is a multi-modal fusion system that creates a holistic "digital twin" for athletes.
Summary updated