Research On Human Fall Detection Based On Deep Learning Algorithm Fusion

According to demographic data released by the National Bureau of Statistics,with the continuous expansion of the elderly group,China has entered an aging society at the present.A series of injuries caused by accidental falls are the second leading cause of accidental death in the elderly.Therefore,the detection of falls in elders living alone can reduce their risk of death and injury.It is of great significance to study how to detect falls for the safety and health of the elderly.In fall detection algorithms based on computer vision,human pose estimation,and convolutional neural networks are widely used.However,the current algorithms still have a lot that can be optimized in feature extraction and network frame size.Aiming at the problems existing in the current fall detection,this thesis carried out the following work:1.Aiming at the problem that some current algorithms use human key points instead of body posture information to detect falls,the error of key points marking will lead to the error of detection results.This thesis proposes a fall detection method based on Openpose and Mobile Net V2.This thesis aims to use the original image information to correct the deviation in the process of key point labeling,improve the effectiveness of feature extraction,and improve the classification accuracy of the algorithm.First,Openpose was used to extract key points of the human body and annotate them in the original image.Then,the improved Mobile Net V2 network is used to extract the features of the original image and marked human posture information to detect falls.Given the problem that some samples in the UR data set have too dark light,which leads to a large deviation in shutdown extraction by point,this thesis brightens relevant data to improve the accuracy of key point annotation.The accuracy of this method on Le2 i and UR data sets is 98.6% and 99.75%,respectively,higher than that of the comparison methods listed.2.Given the problem that the current fall detection algorithm based on Mobile Net V2 cannot take into account both the data background information and the position-coding information,we improved the inverse residual structure of Mobile Net V2.Firstly,we added the channel attention and spatial attention mechanism before point-wise convolution and space convolution,respectively.Then,we improved the output part of the network.The CBAM was added to the beginning part of the classification structure to correct the feature information extracted by the convolutional layer.The above operations improve the network’s attention to the crucial information without increasing the network computation,thus improving the classification performance of the network.Aiming at the problem that there are few fall data in the UR data set,we expanded the relevant data and adopted the methods of translation,shrinkage,and random clipping to expand the fall data to four times the original size,so that the network can fully learn the characteristics of fall movements and detect falls more accurately.The detection accuracy of this algorithm on the common data sets Le2 i and UR is 98.8% and 99.7%,respectively.Although the detection accuracy in the UR data set is slightly lower than that of the first proposed method,the network framework of this method is significantly smaller than that of the first proposed method,and the accuracy is higher than that of the comparison methods listed.