Intelligent Marshland Monitoring System Based On Object Detection And Tracking

In recent years,with the vigorous development of China’s inland water transportation industry,the number of vessels navigating in the inland waterways has increased continuously,resulting in a higher risk of accidents.In order to reduce the risk of vessel navigation accidents,it is urgent to set up an intelligent monitoring system based on machine vision on the river bank,to carry out flow statistics on vessels navigating in the inland waterways,and to ensure a good passage environment.Ship traffic flow statistics based on object detection and tracking have gradually become a research hotspot.However,due to the complexity of the environment in the inland waterways,the phenomenon of mutual occlusion between vessels is severe,and this research is aimed to two problems: the first is the low accuracy of ship detection,and the weak ability of the model to locate small objects in complex environments;the second is the high error rate of ship tracking,and the low robustness of the model in tracking occluded objects.Existing research results cannot meet practical application needs very well.Therefore,this thesis optimizes the object detection and tracking algorithm of ships in complex environments,and designs a practical intelligent marshland monitoring system.The main content of this thesis can be divided into three parts as follows:(1)For ship detection,the RSTS-YOLOv5 object detection model based on the residual sliding window attention mechanism is designed.This model is based on detection model YOLOv5,and improves the model’s ability to extract and learn context information in complex environments by using the residual sliding window attention RSTS(Res Swin Transformers).At the same time,the loss function is changed to Alpha-EIo U loss.The AP50 of the model reaches 94.5%,which is higher than most widely used object detectors.(2)For ship tracking,an improved Deep Sort algorithm for multiple-object tracking is designed.This algorithm improves the original deep feature extraction network using the hierarchical split-residual network HS-Res Net,and enhances the model’s re-identification ability by adding a learning rate warm-up and cosine decay method.Meanwhile,the algorithm improves the trajectory association method in the original Deep Sort algorithm by including low-confidence detection boxes in the matching range and re-matching trajectories that were not successfully matched,thereby improving the tracking ability of the algorithm in complex scenes.The experimental results indicate that the MOTA and MOTP of the algorithm are 65.8% and67.9%,respectively,which has achieved significant improvement compared to the original Deep Sort algorithm.(3)For ship traffic flow statistics,the intelligent marshland monitoring system is designed.This system integrates the ship traffic flow statistics algorithm designed in this thesis,collects real-time image data of inland waterways using a cloud platform spherical camera,and records real-time statistics results using the SQLite relational database.During the system testing phase,statistics were conducted on the ship flow in the Wuhan section of the Yangtze River Basin under two different lighting scenarios,day and night,to verify the practicality and robustness of the system.The results show that the accuracy of ship counting in this system can reach 97.3%,which can meet the application requirements in actual production environments.