Research On High Speed Train Data Dump And Driver Violation Detection System

With the development of the national railway industry,the speed of train operation and the volume of passenger and freight transport are increasing.In order to ensure the safe operation of the train and reduce the human factors causing accidents,a large number of operation records and monitoring video data are generated when the train is running.Due to the huge amount of data,the current train data(mainly video data)dump and analysis is mainly completed by manual operation,which is inefficient,costly and prone to human destruction.Therefore,for ensuring the safe operation of trains,it is very important to realize automatic landing of train data and automatic identification of drivers’ illegal operations.In this thesis,the high-speed dump of train data and driver violation detection are systematically studied.In the data dump stage,an automatic and high-speed vehicle-to-ground data dumping scheme is designed with on-board data collection,pre-storage,and forwarding functions,which greatly improves the data transmission speed by using 5G transmission channel and simplifies on-board antenna installation and reduces cost by using Power over Ethernet technology.In the driver violation detection stage,the improved YOLOv3 object detection model is integrated into LKJ analysis system to achieve automatic detection of driver violation points.The main contents are as follows:Firstly,The high-speed dump scheme for large-capacity data is studied,and the dump host is designed in detail.The main control board uses NXP Qirl Q series processors,configures the interface and memory according to data processing and storage requirements.The POE board uses MK60 chip to receive commands from the main control board,and manage the power on and off of the vehicle antenna.The modular design of the system software,data acquisition,storage and transmission are processed,and the host computer starts to run the system software after power-on initialization.The host will save the data files obtained from the6 A system and CMD system to the local hard disk,and send them to the ground after the train returns to the depot and establishes a connection with the ground.The Ground Data Processing Center can also transmit the data back to the train.Simulate vehicle-to-ground file transmission experiment to verify the function of the scheme.Antenna peak sending rate reaches 1.7Gbps and average rate is above 1.5 Gbps.Secondly,we extract frames from the train surveillance video to make the dataset,and research methods for improving the accuracy of target detection network.Such as,introducing the attention mechanism and spatial pyramid pooling into YOLOv3;modifying the loss function by using DIo U Loss and Focal Loss to speed up the convergence of network training;using Mosaic data enhancement strategy and K-Means++clustering algorithm generates anchor boxes to make the model have a higher m AP value,which is more suitable for this dataset.Our model reaches 0.943 on m AP,which is 0.142 higher than the original YOLOv3 and provides more accurate prediction results for the driver violation item point detection task.Finally,Based on the above research,the high-speed dump system of train data and the driver violation detection system are designed,and the antenna deployment schemes under different requirements are given.Achieve fast dump of large-capacity car data and automatic detection of driver violation points.The system has been successfully applied to Chongqing Locomotive Depot and Nanjing Metro Line Ningju.The onsite dump speed has reached 1.5Gbps,and the accuracy rate of driver violation items detection is 87.9%,basically reaching the expected goal.