Building a skeleton-based 3D body model with angle sensor data

Abstract

Human body activity recognition is beneficial for health monitoring, and wearable angle sensors are one of the most significant techniques to provide accurate body motion information, like joint angles and bone lengths. However, angle sensor data are partial and separated, which is inconvenient to describe a whole body motion. In the paper, we built a skeleton-based 3D body model to recover whole body motions with angle sensor data. We took advantage of two sub models: a multi-output regression model to expand body information and a skeleton recovery model to produce joint positions. The multi-output regression model included Gradient Boosting and Random Forest learning regressions. The skeleton recovery model made use of the law of cosines and binary quadratic equations. In addition, we developed an Android application prototype to show the practical scenarios of the 3D body model. In the evaluation, we measured the performance of the multi-output regression model by calculating statistic metrics, including mean absolute error (MAE), mean square error (MSE) and R2 score. Moreover, we computed the absolute errors and error percentages of the 3D body model for different joints and different motions. In addition, we evaluated the impact of bone lengths to make the evaluation complete. In total, the accuracy of this 3D body model achieves about eighty percent, while most joints’ bias are limited into 10 cm in real world datasets.