Research And Implementation Of Indoor Human Fall Detection Based On Deep Learning

Elderly fall incidents are a serious issue faced by aging societies,affecting the health and quality of life of the elderly,and imposing a heavy burden on families and society.Deep learning technology can learn and train human behavioral data to obtain human feature information.This thesis is based on a project in a practice base in Guangdong and uses deep learning technology to extract key information such as posture and position to detect elderly falls in a timely and accurate manner and provide immediate alerts.Therefore,this work has theoretical and practical application value and significant social benefits.Applying deep learning technology to fall detection has become a current focus in theoretical research and practical applications,with some achievements made in recent years.However,in-depth research and analysis reveal some limitations in existing algorithms,such as insufficient selection of effective features that affect the accuracy of fall detection,using only a single target detection or pose estimation method,which can lead to inaccurate target detection positions or pose estimation detecting empty areas,and high algorithm computation costs.This thesis addresses the above three issues,focusing on solving the first two.This thesis aims to detect falls among the elderly in indoor environments by combining improved target detection and pose estimation algorithms for human fall detection and designing and implementing an indoor human fall detection prototype system.The main work carried out in this thesis is as follows:First,a new indoor fall detection algorithm combining improved YOLOv5 and Open Pose is proposed.Firstly,YOLOv5 is improved by adding a coordinate attention mechanism to adapt to indoor human targets,introducing GSConv lightweight convolution modules,and simplifying the loss function to reduce the number of parameters and computational load.Secondly,deep separable convolutions are used to replace traditional convolutions in Open Pose to address computational load and power consumption issues,and the loss function is replaced.Then,the improved YOLOv5 is used for human target detection,and the human position information is input to Open Pose for keypoint detection and pose extraction to determine if a fall has occurred.Experimental results show that the proposed algorithm improves detection speed while maintaining accuracy.Second,an indoor human fall detection prototype system based on Py Qt5 is designed and implemented.Object-oriented system design theories and methods are used to complete requirement analysis,system design,database design,system implementation,and testing.Requirement analysis clarifies the system’s functional and non-functional requirements.The system architecture is designed according to system design principles,consisting of input source selection,fall detection,and output detection record modules.Sequence diagrams,class diagrams,and flowcharts are used to complete the structural design of each functional module,and the My SQL database is used for database design.The system is implemented using a C/S structure and developed with Python and Py Qt5.Following the testing specifications of GB/T25000.51-2016,the system’s functionality,security,usability,and real-time performance indicators are tested,and the test results meet the system design requirements.Third,fall detection datasets are enriched.Given the limited availability of datasets for fall detection and their single type,a custom indoor human fall detection dataset is created on top of public datasets,containing approximately 9,000 images,laying the foundation for subsequent research.In real life,falls occur in complex environments;therefore,using threedimensional data for pose estimation would be more comprehensive.This project only uses two-dimensional pose estimation for fall detection.In the future,threedimensional pose estimation will be introduced into fall detection.