Fall Detection Algorithm Research Based On Improved YOLOv7

As the percentage of the elderly aged 65 and older in China’s population continues to increase,the health problems of the elderly have also attracted increasing attention,and falls in the elderly are a major factor leading to health damage in the elderly.A system device that can timely,accurately,and effectively detect whether the elderly have fallen is particularly meaningful.Due to the advantages of computer vision based fall detection systems such as fast real-time performance,low cost,and multiple tasks in parallel,the fall detection system made of it has become a key research project for researchers.In this field,the YOLO series of algorithms have good results in fall detection,but among the YOLO series of algorithms,only the YOLOv5 algorithm and the earlier generation of fall detection have been explored in scientific research.There is no research on fall detection using YOLOv6 and YOLOv7 algorithms,and YOLOv6 is committed to industrial needs and does not meet the requirements of fall detection.Therefore,this article proposes a study on a fall detection algorithm based on improved YOLOv7.This system can detect falling targets in videos or images,and can also avoid identifying lying targets in specified areas as falling targets,Finally,it is possible to feed back the fall results detected by the system onto the graphical interface of the system.The main research contents of this article are as follows:(1)Analyze target detection algorithms,compare them from various aspects,collect data through various channels,filter appropriate images,use data to enhance rich samples,and build your own dataset.Mark the fall and stand categories.Use the YOLOv7 network model to train the self-made dataset.With the Io U set to 0.5,the m AP value of the fall and stand targets reaches 86%,It has a good effect in detecting falls.(2)In view of the reduction of key information in the deep feature map and the incomplete extraction of key feature information in the image due to multi-layer convolution operations,this article designs three modules to be embedded in the network model by using attention mechanisms and the idea of expanding convolution,which can give more attention to the key features of falling targets,increase the receptive field,and improve the efficiency of elderly fall detection algorithms.The specific solution for the improvement point is to successively embed three self designed modules on the network structure of YOLOV7,namely,the Attention module(Atten),the Adaptive Average module(AAM),and the Feature Enhancement module(FEEM).The test results of the improved model and the YOLOV7 training model before the improvement are compared.The experimental results show that when the three modules are embedded into YOLOv7 simultaneously,the m AP value of the final trained model for falling and standing targets is increased to 87.6%,greatly improving the detection accuracy.(3)Because it is difficult to distinguish a character between a falling and a lying position,this paper divides the area of the sofa and the bed.When a character is detected lying down or falling in a designated area,it is set to stand,which can effectively avoid misjudgment of the character’s lying down state.(4)Using the improved model to detect falls on video images and displaying the test results on the system interface can better achieve automatic detection of human falls.