The Research Of Sea-surface Target Visual Detection Technology Of USV

In recent years,the sea-surface target detection technology,because of its intelligence and accuracy,has gradually become a hot research area.It is a technology could analysis the video and still images by computer,playing an important role in both military and civilian areas.The ship will be segmented from the background if appear in the image,then their behaviors will be judged,analyzed and tracked.Main purpose of this paper is to set up visual perception system of an Unmanned Surface Vehicle(USV),to check whether there is a boat in the way in order to guarantee safety of USV.In this paper,according to the characteristics of sea-sky images,the target detection technology of sea-surface target can be divided into two types.One is distant ships which appear near sea-sky line;the other is short-range ships which appear near sea-surface.This paper also gives appropriate algorithm for testing on the basis of two different imaging characteristics.Due to the particular sea environment,the collected original images will be vulnerable to the interference of external factors and cannot be used for direct target detection.The correlative technic utilized in the process of sea-sky image preprocessing will be introduced in this paper.Firstly this paper puts forward an improved Hough to detect the sea-sky line and find the starting and ending coordinates of x and y axis;Then get the sea-sky junction area which contains distant targets by translating the Midpoint ordinate of sea-sky line a certain height.Due to the size of distant ship is very small,when sea-sky junction region accounted for one-third of the height of the image could avoid target area separated.The sea area is divided into 20 units using the theory of gray largest variance and statistic the number of pixels between each cell.If there are more than three consecutive numbers of pixels are greater than the set threshold,then,merger the interval and this area is the position where target located.The paper gives the specific method to adjust threshold for different situations.Finally,split out the target with grab cut on the basis of the Pre-positioning of the target.In this paper,the position-near target detection is realized through analyzing sample by means of characteristics of gradient direction histogram(HOG)combined with support vector machine(SVM).Firstly,analyzing the characteristic of the HOG sample and inputting the SVM for training;then analyzing the HOG features of the image to be detected and finally utilizing multi-scale function to detect the target.