Research On Fall Detection For The Elderly At Home Based On Deep Learning

At present,the aging phenomenon in China is becoming more and more serious,and falling is one of the biggest threats to the health of the elderly.When a fall occurs,if not treated immediately,it will seriously harm the physical and mental health of the elderly,and even take away their lives.Therefore,the study of falling detection for the elderly is of great significance to reduce the physical and mental injury of the elderly and reduce the medical cost.Although scholars at home and abroad have done relevant research,due to the diversity of human posture and complexity of behavior,the following problems still exist in the process of fall detection.(1)Due to the non-rigid characteristics of the human body,in daily life,the human body posture will present diversity.For example,the human body will present various postures such as sitting,standing,bending.The image characteristics of human body under camera are also different under different weather and different lights.Some non-human movements(cats and dogs,etc.)will also bring interference to target detection;Therefore,the first key problem is the low precision and recall rate of human targets with different postures due to the diversity of human postures,light changes and false target interference.(2)The behavior of the elderly at home indoors is often accompanied by the following behaviors: walking normally,falling down while walking normally,bending over while walking normally,squatting down while walking normally,suddenly getting up from the seat,bending over while walking,etc.If the elderly live in villas,there will also be behaviors such as going up and down stairs indoors.These behaviors are called falling-like behaviors,which will seriously confuse the falling detection of the elderly.Therefore,the second key problem is that it is difficult to distinguish falls from fall-like actions due to the complexity of human behavior,which leads to misjudgment.In view of the above two key problems,the main work and innovations of this paper are as follows:(1)To solve the problem of low precision of human target detection,we improve the candidate box selection method and Darknet feature extraction network structure of YOLO3 by analyzing human features with different postures.The rich features from human targets with different postures(standing,bending down,etc.)can be extracted,and the precision of target detection can be improved.To solve the problem of low recall rate,we propose a multi-pose human target detection method based on frame difference method and improved YOLO3.The experiments are performed under different illumination,different scenes and different perspectives,and the results show that the improved YOLO3 and fusion frame difference method have strong robustness,and can better improve the detection precision and recall rate of human targets with various postures.(2)It is difficult to distinguish between falling and fall-like behavior,which leads to misjudgment.In this paper,from the perspective of the falling process of the elderly,the complex and diverse behaviors of the elderly are divided into three categories: before falling,during falling and after falling.A fall detection model for the elderly is designed,which is composed of a 10-layer convolution block network.The experiments are done under the conditions of mixed falling behaviors(going upstairs,going downstairs,bending over,squatting,walking)and falling behaviors.The results show that the convolutional neural network based on three-state classification can better distinguish falling and falling behaviors and reduce the occurrence of misjudgment.(3)In order to verify the effectiveness of our algorithm in the real scene,the TensorRT acceleration framework is used and a practical fall monitoring system is designed for the elderly at home based on NVIDIA Jetson Nano artificial intelligence development board.The system consists of edge devices and servers,and has the functions of image preprocessing,model compression,fall display,storage and alarm setting.The system can complete the falling detection task well in the real scene.