Research On Surface Object Detection Technology Of Unmanned Surface Vehicle Based On Deep Learning

With the increase of nautical activities,the demand for monitoring maritime traffic conditions,detecting maritime objects,protecting trade ships,and coastal defense is also increasing.In addition,we also face some problems,such as maritime accident rescue,marine pollution monitoring,marine aquaculture automation and other issues.The above problems are inseparable from the work of the USV,and if the USV can navigate autonomously on the sea or have an operator remotely control the navigation,the USV needs to have the ability to perceive the environment.The object detection technology based on visible light is a must for the USV.An important part of environmental awareness.In this thesis,surface objects(buoys,yachts,container ships,bulk carriers,and cruise ships)are taken as the research object,and a data set of unmanned surface object detection is produced,and a real-time object detection algorithm is studied to solve the problem of real-time detection of unmanned surface objects.It is of great significance to promote the innovation and development of ships and marine engineering equipment for the development space of marine economy.(1)Aiming at the small number of open sea object datasets,low-quality datasets,and many repeated pictures,this thesis considers the needs of the research topic,obtains about 40,000 pictures from the SHIPSPOTTING website,and carefully screens out 2665 pictures that meet the research requirements of this topic.The images were manually labeled using the Label Img tool.For the situation that the same type of object in the picture is similar in size,all pictures are randomly scaled according to the actual situation to improve the richness of the data set.While zooming,generate a new annotation file based on the original annotation file of each image.Use color enhancement in the HSV color space for all images in the dataset to improve the performance of the trained network model.(2)This thesis briefly describes the research status of algorithms in the field of object detection,and introduces the current mainstream research methods.It focuses on the basic principles and network structures of SSD,Faster R-CNN and Yolov4,which are currently three object detection algorithms with high performance.This thesis uses the USV5330 data set and the already public Seaships(7000)data set to train the above three algorithms.Analyze the experimental results from multiple indicators such as loss image,50 m AP and FPS to find out the best performing object detection algorithm and verify the effectiveness of the USV5330 data set.Finally,Yolov4 with the strongest comprehensive performance was selected as the basic algorithm of this research topic.(3)This thesis designs the MLC-CSPP_Yolov4 unmanned surface object detection algorithm based on Yolov4.In this thesis,the MLC module and CMP module are used in the backbone network instead of CSPX to reduce the network depth,strengthen the connection between deep semantics and shallow semantics,and enhance the ability of network feature extraction.This thesis adds the proven effective CSP structure to the SPP network structure to form a new CSPP network to reduce the network weight parameters,making the network easier to train.In the inference stage,the detection speed is improved by merging the branches where the known weight parameters are located to reduce GPU computing consumption.(4)In this thesis,the model trained by the improved object detection algorithm MLCCSPP_Yolov4 on the USV5330 data set is applied to the actual detection system,and two object detection systems for different scenarios are developed,namely the boat-borne detection system and the shore-based detection system.The two systems have been tested separately,and they can run normally,which meets the needs of this research.