Improved Fall Detection Model on GRU Using PoseNet

Abstract

This paper investigate an improved detection method that estimates the acceleration of the head and shoulder key point position and position change using the skeleton key point information extracted using PoseNet from the image obtained from the low-cost 2D RGB camera, and improves the accuracy of fall judgment. This paper propose a fall detection method based on the post-fall characteristics of the post-fall, the speed of changes in the main point of the human body, and the change in the width and height ratio of the body's bounding box. The public data set was used to extract human skeletal features and train deep learning, GRU, and as a result of experiments, this paper find the following feature extraction methods. High classification accuracy can be achieved, and the proposed method showed a 99.8% fall detection success rate more effectively than the conventional method using raw skeletal data.