Water Target Recognition Method And Application For Unmanned Surface Vessel

The unmanned surface vessel(USV)is a new water power in recent years.It can be used in marine emergency rescue,logistics work,water quality monitoring,hydrological surveying,marine environment mapping,water ecological protection and other fields.And with the progress of computer vision technology,the USV equipped with visual intelligent perception system is easier to identify the water target effectively.Therefore,the research on the water target recognition method has the important practical value.Relying on the USV called “ME120”,this thesis mainly studies an artificial intelligence method based on vision sensor and edge computing platform to recognize and track the water target,which must meet the accuracy and real-time requirements.The main works of this thesis are as follows.Firstly,recent development and future perspective of USV are reviewed.And the research status of visual perception technology of USV is reviewed,including the development of object detection technology and object tracking technology based on deep learning.Secondly,the basic theories and common evaluation indexes of object detection algorithm based on deep neural network are briefly introduced.On this basis,the popular object detection algorithms based on deep learning,including Faster R-CNN,SSD,YOLOv3,YOLOv4,are practiced on the selection datasets full of boat images.The performance of these methods is analyzed and the best algorithm is selected.At the same time,the construction of water targets datasets has been carried out.It mainly includes a series of standardized processing,such as feature analysis,frame extraction,duplications elimination,manual filtering,label annotation,data augmentation and so on.Then,according to the water targets datasets,the YOLOv4 algorithm is optimized by network pruning,loss function modifying,blank labels training and preprocessing based on special weather.In the end,the inference speed of the algorithm in this thesis has been greatly improved by 1.78 times,on the almost same Average Precision compared with the original.In order to further acquire more object information,two solutions are carried on.To meet the needs of color detection about ships or navigation buoy,a color extraction algorithm based on Local Contrast Saliency algorithm is proposed.To improve the robustness of system,an object tracking network named Siamese-RPN is introduced,so that the tracker can help object detection network in a cascade way.Finally,based on the above algorithms,the water targets recognition system for USV is constructed,which mainly includes the graphical user interface made by Qt,and the shipboard deployment method based on the development kit named NVIDIA Jetson AGX Xavier.The experiments on the USV show that the system works well in several scenes,and has achieved good results both in the average precision and inference speed.