A Human Motion Prediction Method by Deep Learning Using Temporal PAFs

Abstract

Recently, research and development on automatic operation and factory automation have been actively conducted. It is considered that support robots working in living environments closely related to humans need to recognize and predict human actions and movements from the information obtained using sensors such as cameras from the viewpoint of safety and cooperation. In this paper, we propose a method for predicting human motion using posture information and motion information obtained from video images. In conventional dimensionality reduction using deep learning and encoding and principal component analysis, there is a drawback that it is difficult to know how valid the reduced dimension is for motion prediction. In this paper, we solve this shortcoming by using motion representation and predicting while maintaining the spatial relationship. Experiments show the effectiveness of the proposed method.