Research On Indoor Elderly Fall Recognition Based On Monocular Camera

According to the data of the 7th National Census of China,China’s population will reach 1.41 billion in 2020,and the proportion of people aged 65 and above will reach 13.5%.The aging process of China’s population is accelerating,and home care will become one of the main retirement models in the future,in which abnormal falls are one of the biggest threats to the elderly living alone.The research on prior prevention,detection and evaluation of falls has become one of the hot spots of current research.With the development of intelligent video analysis technology,it is important to study how to quickly and accurately detect the abnormal fall behavior of the elderly living alone in the surveillance video.This thesis proposes a video fall monitoring system solution based on YOLOv5 algorithm,maximizing the use of captured surveillance video to report the identity of the elderly and the fall image to family members and medical institutions in time when the fall occurs through fall judgment and face recognition,facilitating medical institutions to take corresponding rescue measures in time according to the identity of the fallen person.The specific work is as follows:(1)The training strategy of YOLOv5 fall recognition algorithm is improved to transform the fall recognition task into a video sequence target detection behavior analysis task for elderly people living alone.For the sample imbalance problem(the number of samples for normal behavior will be much larger than the number of samples for falls),this thesis adopts a data resampling method to assign different category weights and sample sampling frequencies during training according to the number of labels in different categories.In addition,this thesis also improves the data enhancement method at the input side.Experiments show that the improved training strategy can make the average accuracy of the model improve by 2%.(2)The network structure of YOLOv5 fall recognition algorithm is improved,and for the problem of poor recognition accuracy of the original YOLOv5 s model,this thesis changes the loss function of the original YOLOv5 model,discarding the original CIo U loss function and using the alpha-Io U loss function.The channel attention module is embedded in its neck network,and the network is focused on the target region to be detected by channel weighting.In addition,this thesis also improves the feature fusion of the original YOLOv5 s model by using a bi-directional weighted feature pyramid for higher-level feature fusion.Experiments show that the improved network structure improves the detection performance of the model compared with the original model without adding a large computational cost,and meets the system requirements for detection accuracy and speed.(3)Based on the improved YOLOv5 algorithm in this thesis,a new video fall monitoring system solution is designed and implemented to capture the monitoring video in indoor environment,which can realize the recognition function of fall action and identity information with high detection accuracy and real-time,meeting the requirements of daily use and achieving the expected design requirements.