Research On Indoor Fall Warning Algorithm Based On Real-time Human Pose Estimation

With the aging trend of China's population,health care has become an increasingly prominent social problem.Due to various factors,the elderly are prone to fall in their daily life.If they cannot be prevented and treated in time,serious injuries will be caused to them.Falling has become the number one cause of injury-related death in Chinese people over the age of 65,according to data from the country's disease surveillance system.In addition,China is currently engaged in nursing industry gap of more than 10 million people,it's imperative to achieve the elderly fall prevention nursing through technical means.In this paper,the main algorithms of fall detection at home and abroad are studied.Aiming at solving the problem that can't realize fall warning effectively in indoor environment at present,an indoor elderly fall warning algorithm based on real-time human pose estimation is proposed.In addition,we develop a fall warning platform based on the combination of YOLO V3 pedestrian detection and OpenPose human pose estimation algorithm to solve the problem of indoor environment fall warning.The main work completed is as follows:(1)In order to solve the problems of YOLO v3 in indoor pedestrian detection,such as slow speed,large computation and low detection accuracy,this paper proposes a lightweight YOLO v3 indoor pedestrian detection algorithm based on spatial pyramid pooling.Firstly,in order to improve the speed of detection,the backbone network of YOLO v3 is lightweight processed to delete redundant levels and irrelevant information.Then,according to the unicity of indoor pedestrian detection,the loss function of YOLO v3 is simplified to reduce the calculation amount.Finally,the spatial pyramid pooling is introduced into the YOLO v3 feature extraction network to improve the multi-scale characteristics of indoor human body detection,so as to improve the detection accuracy.Experimental results show that based on the original YOLO v3 algorithm on the self-made indoor pedestrian data set,the improved model can improve the speed of indoor pedestrian detection by 8FPS,and the detection accuracy by 8.4%.(2)With a view to resolving the problems of large power consumption,poor real-time,low detection accuracy of current human pose estimation algorithms based on convolutional neural network,this paper proposes a deep separable convolutional feature extraction network based on lightweight attention mechanism for human pose estimation model OpenPose.Firstly,the backbone network of the original model is improved by replacing the traditional standard convolutional layer with the deep separable convolutional layer to reduce the calculation amount and the number of parameters in the process of feature extraction,so as to achieve the real-time detection effect.Then,a lightweight attention module is introduced into the original model and added behind the deep separable convolution layer.This module improves the accuracy of the extraction of human key points without significantly increasing the amount of calculation.The experimental verification shows that the improved model proposed in this paper has a great improvement in detection speed and accuracy compared with the original model,in which the detection speed is increased from 7FPS to 14 FPS,and the extraction accuracy of human key points is increased from 78.0% to 85.6%.(3)An indoor fall warning platform is designed by combining the improved YOLO V3 human body detection algorithm with the improved OpenPose human body pose estimation algorithm.LSTM mechanism is introduced to store and remember the human pose estimation sequence of the input video,so as to achieve the purpose of predicting the fall.The final experimental results show that the indoor fall warning algorithm designed in this paper can judge the occurrence of fall behavior 300 ms in advance,and the detection accuracy of fall reaches 97%.The algorithm proposed in this paper can carry out fall warning for human bodies in indoor environment by means of human pose estimation,which can predict the occurrence of fall behavior 300 ms in advance,and the detection accuracy of fall reaches 97%,which meets the real-time and accuracy requirements of the warning system.It provides a prototype system for its application in hospitals,health care homes and community service stations.