Human Fall Detection Algorithm Based On Attention Mechanism And Transformer

With the increasingly serious problem of aging population worldwide,falls have become the second leading cause of accidental death in the elderly.However,if fall detections in the elderly are addressed early,the serious consequences of falls can be reduced.With the deepening of research on fall detection in the elderly,many classification algorithms have been developed.However,to better apply them in practical scenarios,the accuracy and real-time performance of most existing fall detection algorithms need to be further improved.Therefore,it is necessary to study fall detection algorithms that have high accuracy while maintaining good execution speed.This thesis conducts research on the fall detection algorithm,and the main works are as follows:(1)To address the issue of low accuracy in current fall detection algorithms,a fall detection algorithm based on adaptive keypoints attention module and time feature extraction is proposed in this thesis.Firstly,a pose estimation network is used to extract human keypoints,and sliding window is used to preprocess the extracted keypoints.Then,an adaptive keypoints attention module is designed to enhance keypoints that are strongly correlated with falls.Finally,the long short-term memory network is improved and used to extract the dynamic time features of the enhanced keypoints,and a fully connected layer is used for fall and normal classification.The experiment is conducted on the UR dataset and Le2 i dataset,and the accuracy of the proposed algorithm on the two datasets is 99.73% and 99.62%,respectively.(2)To address the problem of poor real-time performance of current fall detection algorithms,a fall detection algorithm based on pyramid network and feature fusion is proposed.First,an image reduction module based on convolutional neural network is used to reduce the size of the image.Then,the Transformer is improved and constructed into a pyramid network for extracting spatial features.Finally,a feature fusion module is designed to fuse features from different scales in space,and gated recurrent units are used to extract the temporal features of the fused feature tensor.A fully connected layer is used to classify fall and normal postures.Experiments show that the accuracy of the proposed algorithm on UR and Le2 i datasets reaches 99.61%and 99.33%,respectively,and the execution speed reaches 109 fps.(3)In view of the insufficient ability of convolutional neural network to extract global features and the insufficient ability of Transformer to extract local features,this thesis proposes a fall detection algorithm based on a dual-stream network for extracting both local and global features.The network consists of two branches: one branch composed of a convolutional neural network and a region attention module for extracting local features from the image,and the other branch composed of an improved Transformer for extracting global features.Then,a feature fusion module is used to fuse the local and global features for classification and fall detection.The results show that the proposed algorithm has good performance,achieving accuracies of 99.76% and 99.51% on two public datasets,respectively.