[2D1-GS-2] Machine learning: Evolutionary computation / Network

[2D1-GS-2-04] Motion Discrimination Model from Sensing Point Cloud Data based on One Dimensional CNN and Data Augmentation

〇Kazuchika Suzuki¹, Miho Mizutani¹, Koki Yamada¹, Ayako Yamagiwa¹, Masayuki Goto¹ (1. Waseda University)

Keywords:Action classification, Human motion recognition, Data augmentation, Time-series model, Point cloud data

In recent years, the importance of sensing technology for the safety of residents in elderly care facilities has increased, and in particular, there are growing expectations for the use of time-series point cloud data acquired by millimetre-wave radar. Many methods based on deep learning have been proposed for motion discrimination using point cloud data. Here it is necessary to construct a feature space that contains sufficient information for identification and to prepare a large amount of supervised data. However, the number of points in the point cloud data acquired from millimetre-wave radar, which is the subject of research, varies with each acquisition time. Because of that, these point clouds cannot be used as input to the deep learning model as they are. In addition, there is the problem that the acquisition of data with motion label is costly. In this study, we propose an action recognition method that constructs a feature space using various statistical information over time and models it with a One Dimensional CNN that performs convolution in the time direction, along with preparing sufficient training data through data augmentation method to improve recognition accuracy. Furthermore, we apply the proposed method to actual point cloud data and confirm that the proposed method is capable of motion identification and that the accuracy is improved by data expansion.