Fall Detection Based on Action Structured Method and Cascaded Dilated Graph Convolution Network

The research of fall detection is a hot topic in computer vision. Most existing methods only detect the fall in simple scenes of a single person. Moreover, these methods only extract fall action features from RGB images, and neglect to extract features from human joint coordinates, resulting in a decrease in recognition accuracy. In order to extract discriminative action features, a fall detection method based on action structured method and cascade dilated graph convolution neural network is proposed. The action structured method (ASM) is proposed to model the skeleton of human action through the pose estimation algorithm, which removes the interference of complex background. Besides, the object detection algorithm is utilized to locate multiple people to transfers the fall detection issue of multi-person to single person fall detection. The proposed cascaded dilated graph convolution network (CD-GCN) enlarges the receptive field by the dilated operation, effectively extracts action features from joint node coordinates, and fuses multichannel features with different dilation rates, then finally obtains the classification results. The proposed method achieves the best accuracy on three public datasets and one self-collected dataset, which is out-performing other state-of-art fall detection methods.