Behavior Estimation System for Elderly Individuals based on 3D LiDAR and Self-Learning Function

In recent years, population is aging rapidly in Japan, and the shortage of caregivers in the welfare field is a serious problem. Therefore, a monitoring system is being considered to observe the lives of elderly individuals in group homes on behalf of their family members living in the remote location. In the existing studies, the system for estimating human behaviors using cameras and handheld sensors are proposed and developed. However, these systems have problems related with invasion of privacy and discomfort when the user carries the sensor. Therefore, in this study, we propose a new system to observe the behavior of residents by utilizing 3D LiDAR (Light Detection and Ranging) that can accurately measure the three-dimensional structure of target objects as point clouds. The proposed system estimates positions and postures of residents by applying a deep learning technology to point clouds measured using the 3D LiDAR. In addition, generation of training data for building machine learning models requires an annotation process on 3D point clouds for setting a correct label to each part of the point clouds, which takes significant time and effort. Therefore, in this study, we propose a method for supporting to generate training data by automatically setting the correct labels to appropriate parts of the point clouds by linking image analysis and machine learning technologies.