Hybrid Sensing: Redefining 3D Posture Estimation for Lower Back Healthcare

TL;DR

Low back pain (LBP) is a global health crisis, often exacerbated by improper posture during unsupervised rehabilitation. This paper presents a dual-IMU wearable system combined with computer vision (MediaPipe) to track spinal and wrist movements in real-time. By fusing sensor data with 33-keypoint skeletal tracking, the system provides a high-fidelity 3D estimation of a user's posture, achieving up to 98% accuracy in dynamic movement classification.

Problem & Motivation: The Gap in Home Rehabilitation

Physical therapy is the gold standard for LBP, but it faces a catch-22: clinical supervision is expensive and time-consuming, while home exercise often lacks the proprioceptive feedback necessary to prevent further injury.

Existing solutions typically suffer from two flaws:

1. Limited Scope: Many wearables only monitor the spine, ignoring hand/arm movements that affect overall balance and spinal load.
2. Lack of Context: Vision-only systems can be obscured by clothing or environment, while sensor-only systems struggle with global spatial orientation.

The authors' insight was to create a synergistic "Sensing + Vision" pipeline that offsets the drift of IMU sensors with the absolute spatial references provided by computer vision.

Methodology: The Fusion of Silicon and Pixels

The system architecture is divided into three distinct pillars:

1. Dual-IMU Hardware Layout

Unlike single-sensor setups, this system uses two nodes:

• Back Device: An Arduino Nano BLE Sense placed at the lumbar region to track spinal tilt and rotation.
• Wrist Watch: An Arduino Nano 33 IoT that tracks arm movement, which is critical for exercises like squats or bird-dog poses.

2. Computer Vision (BlazePose)

The researchers utilized Google’s MediaPipe/BlazePose framework. Unlike older YOLO models, BlazePose is optimized for human pose estimation, tracking 33 keypoints. In this study, the authors focused on the lumbar-pelvic "hinge" points (11, 12, 23, 24, 25, 26) to quantify the quality of exercises.

Fig 1: The information flow from dual-wearables and vision streams to the forecasting model.

3. Mathematical Alignment

To bridge the raw sensor data and 3D space, the system employs Quaternions ($a+bi+cj+dk$). This avoids "Gimbal Lock" and allows for a more stable calculation of the body’s rotation angle $\alpha$ relative to the gravity vector.

Experimental Results: Precision in Motion

The system was tested against various rehabilitation poses, including high planks, glute bridges, and squats.

• Dynamic Accuracy: The system excels at tracking motion. Time-series forecasting (ARIMA) yielded accuracies of 96.7% to 98.7% for active lateral movements.
• Vision-Based Validation: The AI exercise tracker successfully monitored squat completion by calculating the vertex angle at the hip, identifying the transition from 180° (standing) to 77° (peak squat).

Fig 2: Real-time pose estimation using BlazePose during various therapeutic exercises.

The Standing Paradox

Interestingly, the system showed a slight drop in accuracy during static "Standing" or "Tilted" positions (~84-89%). This is a common phenomenon in IMU-based sensing where the lack of signal variance (acceleration) makes "stillness" harder to distinguish from "sensor noise."

Critical Analysis & Conclusion

Takeaway

The core value of this work is the scalability of the localized sensing. By moving beyond a single lumbar sensor and integrating cost-effective vision (even a laptop camera), the authors have created a framework that provides clinical-grade feedback at a consumer price point.

Limitations & Future Work

• Environment Sensitivity: The vision component still relies on lighting and camera placement.
• Deep Learning Integration: While ARIMA was used for preliminary forecasting, the authors suggest that a Deep Neural Network (DNN) would better handle the non-linearities of complex skeletal movements.

The future of LBP management likely lies in this "Smart-Body" approach, where our clothes and our cameras work together to keep our spines aligned.

Keywords: Posture Estimation, IMU, MediaPipe, Low Back Pain, IoT Healthcare.