The Visual Accelerometer: A High-fidelity Optic-to- Inertial Transformation Framework for Wearable Health Computing

Human activities of daily life (ADL) monitoring has been applied in life-critical applications such as occupational safety and stroke rehabilitation tracking. However, wearable computing, as the main technical paradigm of ADL monitoring, requires tremendous efforts to obtain satisfactory inertial data with labels for training. In this paper, we develop VisuaIAcc, a high-fidelity optic-to-inertia framework of human locomotion for wearable computing, which leverages harvested light-intensity data from public videos to reconstruct authentic wearable motion data. Specifically, a two-step optical motion estimator is first designed to infer the high-quality optical motion field (OMF) from the time-varying light intensity. Then, the obtained OMF is fed to an optic-to-inertia transformer, which leverages human kinematics constraints in light ray projection to recover time-sequential inertial data in a convolution-based process. Experimental results show over 0.86 Pearson Correlation Coefficient between reconstructed data via VisualAcc and ground truth from authentic off-the-shelf MEMS sensors. Furthermore, we conduct a case study on IMU inverse human dynamics analysis to show VisualAcc's potential in empowering and transforming fine-grained wearable computing.