착용형 로봇의 선행제어를 위한 하지 관절운동의 동작분할 패턴 분석

초록

Objective: This study analyzes the angular variation patterns of clustered 3D skeletal data through unsupervised learning to characterize lower extremity joint motion activities and propose a motion segmentation system. Background: For the commercialization of wearable robots, it is important to recognize and react to the user's intentions in time so that they can move appropriately according to the operator's joint movements. To improve human intent prediction performance, computers should apply motion segmentation techniques to efficiently learn human behavior. Method: Angular data for the back, hip, knee, and ankle joints are extracted from 3D skeletal data using a kinect sensor for 6 major lower extremity working activities of 4 male subjects. It compresses high-dimensional data through a CNN-based autoencoder and analyzes joint movement patterns between clusters by performing KMeans Clustering. Results: Unsupervised learning on motion patterns showed that there is a clear pattern between clusters for the 6 major working activities. the motion segmentation of the lower extremity joints is classified into clusters with patterns of back (3), hip (4), knee (3), and ankle (3). Conclusion: The combination of clusters provides a simple representation of the 6 major working activities and can be utilized as an approach to represent more complex activities. Application: Beyond simply classifying motion patterns, it is expected to be used in the development process of algorithms for motion prediction.

키워드