Research On Fall Detection Algorithm Based Improved YOLOv7

Due to the increasing number of elderly people,the aging problem of Chinese population is becoming more and more serious,and the health problem of old people has been paid much attention.According to the investigation,fall is one of the main reasons affecting the health of the elderly.If the elderly living alone can get help and treatment in a short time after fall,their physical safety can be guaranteed in the first time.Therefore,it is of great significance to study the fall detection of the elderly.This paper focuses on a fall detection method based on computer vision.Since YOLOv7 is one of the fastest and more accurate target detection algorithms,this paper proposes a lighter and more accurate fall detection network model based on the YOLOv7 target detection algorithm to address the difficulties of redundant background information in fall detection tasks and the large computational effort of the YOLOv7 algorithm that affects the detection speed.The specific improvements are as follows:(1)To address the problem of interference from background factors,an attention mechanism module is added to solve the problem.In the fall detection algorithm,by introducing the attention mechanism,it makes the algorithm focus more on the fall detection part of the character,obtain more detailed and relevant information,and ignore irrelevant background information,so as to improve the detection effect of the target of concern and achieve the purpose of improving the overall detection effect of the network model.(2)To address the problem of ambiguous aspect ratio performance of YOLOv7 loss function,the original CloU loss function is replaced by SIoU loss function,the vector angle between the required regressions is fully considered,and the distance is redescribed using vector pinch angle to reduce the loss function degrees of freedom,thus effectively improving the speed of model training and the accuracy of inference.(3)The YOLOv7 network model is lightened and improved to address the problem that the large computational load of the model affects the detection speed.The ELAN and ELAN-W modules in the base network model are improved by combining the depth-separable convolution with the FReLU activation function to form a new convolution module,which ensures the effective improvement of the detection accuracy while reducing the computational effort.The experimental results show that the improved YOLOv7 model reduces 42.1%in floating point operations(FLOPs),improves from 92.1%to 93.3%in mean average precision(mAP),and improves from 87 FPS before to 91 FPS in detection speed compared to the base YOLOv7 model,indicating that the improved network model can be better applied to elderly fall detection task.