Research And Application Of Human Fall Detection Technology In Indoor Scenes

With the growing elderly population in China and increasing severity of aging issues,ensuring the safety of the elderly has become a key topic of concern.Falls remain a primary threat to the health of the elderly,causing irreversible harm and imposing significant burdens on both their physical and psychological well-being,as well as their families.Traditional fall detection algorithms rely on sensors to collect human body data and environmental information,which can be cumbersome to wear,expensive,and difficult to install.In recent years,with the rapid development of deep learning,pose estimation has achieved success in various fields,leading to an increase in pose estimation-based fall detection methods as the mainstream approach.To address the shortcomings of current pose estimation-based fall detection algorithms,such as imprecise extraction of fall features,confusion between falls and similar actions,and incorrect human body posture depiction,this study focuses on human fall detection technology in indoor scenes.Specifically,this research examines adaptive heatmap regression-based and composite field-based pose estimation methods,and attention-enhanced graph convolutional LSTM-based fall detection methods.This thesis presents four main contributions to address these challenges.(1)To address the issue of reduced accuracy in current bottom-up pose estimation methods,which use identical labeling to supervise all key points and cannot accurately detect poses for people of different scales,this thesis proposes an adaptive heatmap regressionbased pose estimation method.This method addresses the ambiguity issues associated with multi-scale human body differences and labeling by adapting each key point’s Gaussian kernel standard deviation through a scale-adaptive heatmap regression approach and introducing weight-adaptive heatmap regression to alleviate the severe imbalance between foreground and background samples.By doing so,the proposed method achieves significant improvements in accuracy compared to other state-of-the-art approaches.(2)To address the problem that multi-person pose estimation methods in dense crowd scenes cannot group key points correctly and lead to incorrect key point connections.In this thesis,we propose a composite field-based pose estimation method,which can accurately detect and group key points through spatial accuracy regression and joint scale estimation,and can connect key points correctly.Comparative tests are done in dense crowd scenes to prove the effectiveness of the method.(3)This thesis proposes an attention-enhanced graph convolutional LSTM-based fall detection algorithm that addresses the confusion between falls and similar actions in existing fall detection algorithms and their inability to extract rich fall features.The proposed method leverages human body posture obtained from(1)and(2)to extract fall features from both temporal and spatial dimensions,integrate key joint information,and classification models.The results of comparative experiments conducted in indoor scenes demonstrate the proposed method’s high accuracy and precision in detecting human falling behaviors.This research represents a significant contribution to improving the accuracy and reliability of fall detection methods and has promising applications in various fields.(4)This thesis introduces a fall detection system on a robot platform that builds on the attention-enhanced graph convolutional LSTM-based fall detection algorithm.The system integrates the robot platform and an Android app through ROS to perform map building,navigation,and fall detection tasks in indoor scenes.By leveraging the proposed algorithm’s capabilities,the system achieves high accuracy in detecting human falling behaviors.This research represents a significant contribution to the development of practical fall detection systems and has promising applications in various fields,such as elderly care and medical assistance.The experimental results demonstrate the effectiveness and reliability of the proposed system,indicating its potential for real-world deployment.