Article Preview
TopA vision-based fall detection system acquires images through video equipment, detects and uses images to detect a fall. Deleting the background from the image (Abobakr et al., 2018; Chen et al., 2020) by extracting and characterizing the human skeleton data from the video, it detects a fall by characterizing the shape or silhouette of a person in the video (Abobakr et al., 2018; Kingma & Ba, 2015). Recently, the method of recognizing human activities based on skeletal data has focused on detecting falls (Lie et al., 2018). Skeleton data is extracted using a 3D depth camera such as Kinect developed by Microsoft or 2D RGB video through CNN-based technology such as PoseNet (Adhikari et al., 2019) In other words, PoseNet can conveniently and quickly extract skeleton data from images captured by a camera. Based on the skeleton information of the human body, it has not only high accuracy, but also simple and inexpensive features (Chen et al., 2020). The extracted skeleton data is spatial-temporal data that changes with time. Recurrent Neural Network (RNN) is a powerful method of processing time-series data classification, but it takes a long time to learn and a part of the result is a problem in which the weight disappears (vanishing gradient problem) or an exploding gradient problem occurs. To solve this problem, the use of LSTM (Long-short term memory) and GRU (gated recurrent unit) was introduced (Chen et al., 2020)(Solbach & Tsotsos, 2017). LSTM does not have the same problem as RNN.
In general, 2D RGB cameras are installed and used by CCTV in hospital rooms and industrial sites of medical facilities. There may be concerns about the protection of some privacy, but it is believed that the camera will not be regulated by the security of privacy regulation because the purpose is to extract the skeleton of the body. GRU has the advantage of short learning time due to its simple structure as a modification of LSTM.
This propose a Fall Detection Method Based on Skeleton Key points Group(FDSG) that classifies and infers human poses from the extracted skeleton data from PoseNet to detects falls pose using artificial intelligence, GRU.