Research On Ferry Vehicle Detection And Statistics Based On Deep Learning

Ferry shipping is an important part of my country’s transportation.The amount of port trade in my country has increased year by year,and the safety of ferry boats has become a top priority.In order to ensure the safety of the ferry during navigation,the ferry administrator often adds a camera at the entrance of the ferry to monitor the boarding vehicles and conduct statistical analysis to provide conditions for subsequent functions such as judging overload,so as to ensure the safety of the ferry.Traditional camera monitoring can only monitor the images of boarding vehicles,and requires naked eyes to distinguish vehicle information,which cannot manage boarding vehicles more efficiently.To solve this problem,this thesis proposes a vehicle statistics method based on detection and tracking.By counting vehicles in real time,ferry managers can monitor the traffic flow on board in time.It helps to ensure that ferries are not overloaded and that scheduling and management decisions can be made accordingly,enabling automated management.Existing target detection algorithms rely on prior anchor boxes,and obtain hyperparameters related to anchor boxes for different sample datasets to determine target candidate regions.The model training operation of this method is complex,the efficiency is low,and it is prone to missed or repeated detection problems caused by unreasonable anchor frame settings.For this reason,this thesis selects the anchor-free YOLOX target detector and Deepsort target tracker as the backbone,and improves the performance of the algorithm through optimization.The main research contents are as follows:(1)Aiming at the real-time detectability problem of ferry entrance vehicle management,this thesis selects YOLOX-S target detection algorithm,and optimizes the backbone network part and loss function part of YOLOX-S target detection.Relevant improvements include: Aiming at the real-time problem of ferry entrance monitoring,select and improve the high-performance residual structure MBConv and Fused-MBConv to replace the residual structure in the original backbone network.In view of the complex changes in the environment at the entrance of the ferry,the ECA channel attention mechanism module is introduced,which can adaptively learn the weight of each channel,and has strong adaptability and generalization ability.Aiming at the problem of false detection of occlusion between vehicles at the entrance of the ferry,the ICIOU loss function is used to replace the IOU loss function used in the original network to ensure the accuracy of detection.(2)To accurately locate the target vehicle in the ferry entrance video and ensure that the ID assigned to each vehicle remains unchanged,this thesis uses Deepsort target tracking technology to track the target vehicle in the video.The algorithm extracts the target features again through the feature re-extraction model,predicts and correlates the target information of the previous frame,the current frame and the next frame through the Kalman filter and the Hungarian matching algorithm.(3)Aiming at the problem of calculating the number of boarding vehicles at the ferry,this thesis uses the baseline method to calculate the number of boarding vehicles.A baseline is set at the entrance of the ferry.When a vehicle touches the baseline,the number of boarding vehicles increases by 1.And by counting the virtual detection line with different position coordinates,find out the most suitable virtual detection line position for vehicle tracking and counting.Establish a ferry management monitoring system and set up an interactive interface with users.In addition to the ferry monitoring module,the ferry management monitoring system also adds a user management module,a ship management module,and an abnormal event a1 larm module to jointly ensure the safety of the ferry.