Skeleton-based fall detection using Computer vision

Abstract

Falling is one of the biggest public health issues that can cause many serious long-term repercussions for patients and their families. In this thesis, we propose an appropriate model for fall detection using graph convolutional networks. Recently, most problems related to human action recognition, including fall detection, can be handled by applying the Spatial Temporal Graph Convolutional Networks model (ST-GCN) using 2D or 3D skeletal data. We take advantage of the transfer learning technique from the NTU RGB+D consisting of 60 daily actions to extract features for the fall detection task efficiently. Besides, to highlight critical frames in the original sequence, we suggest using a temporal attention module. This module consists of two parts: (1) average global pooling, and (2) two fully connected layers to generate an attention score for each frame. We perform experiments on two datasets, i.e., FallFree and TST v2. This leads to a 3.12% increase in the TST dataset and a 2.67% improvement in the FallFree dataset. Notably, with respect to FallFree, the accuracy of the model is up to 100%.