Research On Intelligent Video Monitoring Technology Of Metro Stations Based On Deep Learning

With the rapid progress of urbanization,the construction of smart stations has become the development trend of urban rail transportation.At present,due to the complicated manual workflow and huge workload in metro stations,it is easy for staff to be negligent in crowd control and passenger escalator fall supervision related aspects.In this thesis,we propose PC-YOLO(People Counting YOLO)algorithm and FD-YOLO(Fall Detection YOLO)algorithm for people flow detection and escalator passenger fall detection,respectively,using a deep learning-based object detection method,and deploy the algorithms into a Jetson Nano development board.The main work is as follows:1.The crowd detection dataset Subway Crowd ABPlus and the fall detection dataset Fall Detection Data Plus are constructed respectively.The Subway Crowd ABPlus dataset is compiled from public datasets and self-taken and labeled datasets,with a total of 5698 images containing multiple angles of head objects and can be categorized into sparse,medium and dense crowds based on the number of heads.The Fall Detection Data Plus dataset is a combination of public datasets and self-collected and labeled datasets,with a total of 8925 images of pedestrian falls from multiple scenes.2.The PC-YOLO network-based algorithm for detecting pedestrian flow on metro platforms is designed and implemented.The PC-YOLO network consists of backbone feature extraction network,feature fusion network and detection head;an extra small detection head of size 160*160*64 is added to the network for optimisation;the SIo U Loss function is used as the border loss function.The experimental results show that the detection accuracy of the PC-YOLO algorithm is 92.3%,which is a 2.4% improvement compared to 89.9% for the YOLOv7-tiny network and a 2.3% improvement compared to 90% for the YOLOv5 s network,and that the algorithm has improved detection accuracy in different crowd densities.3.The FD-YOLO network-based algorithm for metro escalator fall detection is designed and implemented.The FD-YOLO network consists of backbone feature extraction network,feature fusion network and detection head;a new C3 AM module is formed by adding the Sim AM attention mechanism to the C3 module of the backbone feature extraction network;the SIo U Loss function is also used as the border loss function.The experimental results show that the detection accuracy of the FD-YOLO model is 85.3% when detection is performed on the test set,which is a 1.8% improvement compared to 83.5% for the YOLOv5 s network and a 3.1% improvement compared to 82.2% for the YOLOv7-tiny network.4.The PC-YOLO algorithm and FD-YOLO algorithm are deployed into the Jetson Nano in order to apply the algorithms designed above to real-world scenarios.The Hikvision DS-2CD3T47WD-LU camera is selected as the image acquisition device,and a Jetson Nano-based detection system is built and optimised and accelerated using Tensor RT.The experimental results show that after the original pt file is accelerated by Tensor RT with FP16 accuracy,the detection speed before and after acceleration of PC-YOLO algorithm is 6FPS and 12 FPS respectively,and the detection accuracy is 92.3% and 91.7% respectively;the detection speed before and after acceleration of FD-YOLO algorithm is 6FPS and 11 FPS respectively,and the detection accuracy is 85.3% and 84.2% respectively.