Research On Ship Target Detection And Tracking Algorithms Based On AE-YOLOv3

Inland waterway shipping has the characteristics of many types of ships,high density of ships,fluctuating with waves,influence of water mist,and complex quay crane environment.To ensure the safe and efficient operation of inland waterway shipping,an advanced and intelligent shipping management system must be established.and the rapid and accurate detection and effective tracking of ship targets are the basis and key technology of the intelligent shipping management system,Its research is of great significance to improve the intelligent navigation of ships and the management level of management departments.Traditional ship target detection algorithms have poor robustness,low detection accuracy and time-consuming on complex inland waterways,which can not meet the needs of ship intelligent navigation management.In recent years,as deep learning has played a brilliant role in various competitions,more and more researchers use deep learning algorithm to detect and recognize ships.Aiming at the problems of traditional ship target detection algorithms,this paper designs the YOLOv3 algorithm based on feature attention and feature fusion enhancement technology as the ship target detector,and the improved Deep SORT tracking algorithm is used as the ship tracking module,the two are combined to realize the automatic detection and tracking of inland ships.The main work of this paper can be summarized as follows:1.Make a ship data set by selecting a certain number of frames from the inland waterway monitoring video,and labeling it as VOC format through Label Img,considering the over-fitting problem,and the number of data sets are increased by data augmentation.Then,based on the surveillance video,the traditional ship target detection algorithm experiment was carried out,and the limitations of the traditional algorithm in inland ship detection were analyzed.2.On the self-built data set,comparative experiments were conducted based on Faster R-CNN,SSD and YOLOv3 algorithms,and the average precision,precision of each category,precision rate,recall rate and time of the three algorithms for ship target detection were compared and analyzed,select the YOLOv3 algorithm with the best performance for optimization and improvement.this paper designs Feature Attention Feature Enhancement YOLOv3（AE-YOLOv3）ship target detector based on feature attention and feature fusion enhancement technology.The feature attention module is constructed by introducing the attention mechanism,which is embedded in Darknet-53 to recalibrate the feature channel,so as to improve the feature extraction ability of the model under the complex navigation background;for the YOLOv3 feature fusion process,low-level feature semantic information insufficient problem,the feature enhancement module is constructed and applied to the feature fusion part to enhance the receptive field size of the corresponding feature layer and the correlation degree of the feature extraction network.Experiments on the expanded ship data sets show that AE-YOLOv3 detector has better results in terms of detection performance and time.3.Combining the AE-YOLOv3 ship target detector with the improved Deep SORT tracking algorithm to realize the automatic detection and tracking of ships.The AE-YOLOv3 target detector is used to detect the surface ships,and then the improved Deep SORT algorithm is used to track the ships.The experimental results show that the system designed in this paper can better detect the category of ship,and can accurately track the corresponding targets.