Design And Implementation Of Human Fall Detection System Based On Skeleton Information

Falling is a common and dangerous behavior in daily life.In order to reduce the harm caused by falling,the identification and detection of falling behavior has become an important research topic in the field of intelligent video surveillance.With the popularization of camera monitoring equipment and the development of artificial intelligence technology,fall detection based on computer vision is a hot research direction at present.Among them,the detection model based on skeleton information has the advantages of small amount of data and less susceptible to complex background interference.It shows good robustness in detection.However,when the detection model is applied to fall detection,further research is still needed on the extraction of skeleton sequences for multiple human targets,and the full use of human falling behavior features to improve the detection accuracy.Therefore,on the basis of analyzing the characteristics of fall behavior,this paper proposes a human fall detection algorithm based on skeleton information.The algorithm includes two parts: skeleton sequence data collection and fall behavior detection,and uses Flask technology to deploy the method on the cloud server,and a fall detection system with BS architecture is implemented.The main contents of this paper are as follows:(1)A multi-body target-oriented skeleton sequence extraction method is proposed,which detects and tracks all the bodies in the actual scene,and separates the skeleton sequence data of each body.First,use the target detection algorithm VFNet to identify and locate the human body,then use the multi-target tracking algorithm Deep SORT to continuously track the human body,and use the multiperson pose estimation algorithm HRNet to extract the human skeleton during the tracking period.Finally,the human skeleton sequence data can be collected through continuous tracking.For the selection of algorithms at each stage in the process of skeleton sequence extraction,different network models are used for comparative experiments.The experimental results show that VFNet,HRNet and Deep SORT are the optimal algorithm models in their respective stages.(2)A subgraph weighted adaptive graph convolutional network SI-AGCN based on an improved skeleton partition strategy is proposed,which takes skeleton sequence data as input to make fall behavior judgments.In order to reduce the confusion of falls and daily behavior,the division strategy of skeleton graph is improved,a three-partition spatial configuration strategy is proposed,and the subgraph weighted adaptive graph convolution is used to replace the spatial graph convolution in the original spatiotemporal graph convolution ST-GCN.part,which constitutes the SI-AGCN network of this paper.Finally,based on three public fall behavior datasets,the Fall-skeleton Dataset is extracted and constructed using the pose estimation algorithm,and training and validation experiments are performed on this dataset.The experimental results show that compared with Compared with the spatiotemporal graph convolutional network ST-GCN,the fall detection accuracy of the SI-AGCN network increased by 6.9% to 96.4%.(3)Flask technology is used to deploy the human fall detection algorithm in this paper on the cloud server,and a fall detection system with BS architecture is implemented.The fall detection system is divided into three parts: image acquisition terminal,cloud server and web browser.First,the Alibaba Cloud ECS server is used,and then Nginx,u WSGI and Flask technologies are used to complete the deployment of the fall detection algorithm.Finally,the deployed fall detection system is tested in the actual scene.The test environment is single and multi-person activities under three lighting conditions.The test results show that the human fall detection system in this paper can meet the requirements of fall detection.