Design Of An Embedded Intelligent Video Monitoring System For Home-based Elderly

As China’s economy continues to prosper and develop,a large number of young people go out to work,and there are many elderly people left alone in rural areas without children around to take care of them,the safety of the elderly has become a difficult problem for their children.To address this problem,this paper designs an embedded-based intelligent video surveillance system for the elderly at home,deploying an improved YOLOv5 s model in the embedded device to achieve real-time detection,while the system implements alarm functions for abnormal events such as elderly fall,fire and pedestrian detection,and supports live broadcast and playback functions,which can better safeguard the lives of the elderly at home.The paper firstly compares the traditional target detection algorithm with the YOLOv5 algorithm by extracting moving human target cases,and selects the YOLOv5 algorithm with better performance and robustness as the system target detection solution.Secondly,a dataset containing 37,684 samples was obtained by collecting fall,pedestrian,flame and smoke samples through self-timer and network search and performing data enhancement,which was used as the baseline data after the training of YOLOv5 s model.Subsequently,the YOLOv5 s model was combined with Mobile Net V3,Ghost and Bi FPN networks for lighter and improved ablation experiments,while combining the more efficient H-Swish activation function in Mobile Net V3,the less computationally intensive Ghost convolution module and the Bi FPN feature fusion network which can better capture feature information at different scales The HSwish-YOLOv5s-Ghost-Bi FPN model is proposed and comparison experiments are conducted.To verify the practical effectiveness of the improved model for deployment on embedded platforms,the Jetson Nano 2GB embedded platform was chosen as the vehicle for deployment.After testing,the detection accuracy and recall of the improved model were comparable to the original YOLOv5 s model under the condition of using Tensor RT acceleration,with m AP 0.5 of 87.1%,recall of 81.5%,average detection time and inference time of 33.6ms and 25.1ms respectively,10.2% and 11.3% faster than the original YOLOv5 s model,and FPS up to 29.8 frames per second,while the model size is reduced by 21.1% to meet the real-time detection requirements,indicating that the improved model is more suitable for deployment in embedded devices.Finally,the thesis combines the improved model,RTMP,RTSP,FFmpeg,ZLMedia Kit streaming server,Py Qt5,FLASK framework,Android and other technologies to implement an end-to-end(embedded end-server end-mobile phone end)embedded intelligent video surveillance for the elderly at home system.The experiments show that the system designed in this paper can meet the requirements of use and has practical application value.