Deep Learning of Biomechanical Dynamics With Spatial Variability for Lifestyle Management

Human Physical Activity Detection (PAD) is essential for advanced lifestyle management. Nowadays, many devices like smartwatches and phones have motion sensors. And the placement of devices, especially phones, is diverse. Understanding the optimal sensor location is important for robust PAD applications. However, the biomechanical dynamics are highly complex, making signal processing very challenging. In this study, we propose to leverage deep learning to perform intelligent PAD, and comprehensively compare seven different body locations for optimal sensor placement recommendation. Multi-stage deep learning has been proposed and developed, which firstly leverages convolutional layers to abstract spatial features from the motion data, and then generates the physical activity type prediction with fully connected layers. Evaluated on the real-world database, the proposed deep learning model is very effective on PAD tasks, and successfully determines the thigh location as the optimal sensor placement method. This study will greatly advance deep learning-driven biomechanical dynamics mining for advanced lifestyle management.