Fall Detection Method Based on Pose Estimation Using GRU

Abstract

Falls are a major cause of injuries or deaths in the elderly over the age of 65 and a factor in social costs. Various detection techniques have been introduced, but the existing sensor base fall detector devices are still ineffective due to user inconvenience, response time, and limited hardware resources. However, since RNN (Recurrent Neural Network) provides excellent accuracy in the problem of analyzing sequential inputs, this paper proposes a fall detection method based on the skeleton data obtained from 2D RGB CCTV cameras. In particular, we proposed a feature extraction and classification method to improve the accuracy of fall detection using GRU. Experiments were conducted through public datasets (SDUFall) to find feature-extraction methods that can achieve high classification accuracy. As a result of various experiments to find a feature extraction method that can achieve high classification accuracy, the proposed method is more effective in detecting falls than unprocessed raw skeletal data which are not processed anything. © 2021, The Author(s), under exclusive license to Springer Nature Switzerland AG.