The Design And Implementation Of YOLOv3-based Fall Detection System For The Elderly In Home

Since the 21 st century,China’s birth rate has been decreasing year by year as well as the increasing number of elderly people,the scarcity of resources in elderly care institutions is serious,and home care will become an important way.Big data shows that sudden illnesses and abnormal falls are one of the main causes of casualties among elderly people living alone at home,and the detection of abnormal falls among elderly people living alone at home under home monitoring scenarios is one of the current video analysis research hotspots.The thesis designs a fall detection system based on the YOLOv3 target detection algorithm,focusing on the application of target detection in the application scenario of home care services.By capturing videos of the elderly’s daily activities,the system identifies the elderly’s fall behaviour in real time,once an abnormality is detected,instantly sends the elderly’s location and abnormal fall information to both the guardian and community health care workers,so as to provide timely and proactive rescue and reduce the injuries caused by the elderly’s fall.The YOLOv3 target detection algorithm has a certain advantage over traditional target detection algorithms and several other mainstream deep learning-based target detection algorithms in terms of detection rate,so this thesis improves the network architecture of YOLOv3.The experimental verification shows the system performance met the expected design requirements.The specific work is as follows:(1)Introduce the background and importance of elderly fall detection research,the current state of research in the country and abroad,and learn about traditional target detection algorithms and the related theory and network structure of deep learning,the comparison and analysis of several mainstream network structures based on deep learning,and the proposed improved algorithm of the YOLOv3 architecture based on deep learning neural networks in this thesis.(2)Based on the network structure of the YOLOv3 target detection algorithm,the BN(Batch Normalization)layer is merged into the preceding convolutional layer as a way to improve the speed of forward inference of the fall detection model.First,the BN layer in the test phase is equivalently replaced with a convolutional layer,and then the BN layer is fused with its preceding convolutional layer to obtain the new weight matrix and bias term of the fused convolutional layer.At this point,the BN layer of the network can be removed and only the convolutional layer can be used,and the new adjusted weights and bias terms can be used to train and test the labelled sample dataset by changing the attribute profile of the network structure.Experiments show that the fall detection accuracy of the improved YOLOv3 algorithm remains largely unchanged,but improves the inference time before and after the model,significantly increasing the detection speed and providing good robustness.(3)The loss function of the YOLOv3 target detection algorithm is improved.In the loss function of the YOLOv3 target detection algorithm,Io U(Intersection over Union)is used to calculate the position information loss function.Since Io U cannot directly optimise the problem that there is no overlapping part between the prediction frame and the real frame when calculating the position information loss function,the thesis uses GIo U(Generalized Intersection over Union)is used instead,and a better detection effect is achieved.