Forward-looking Sonar Underwater Target Detection And Tracking Technology Based On Deep Learning

With the deepening of people's attention to marine resources,the demand for exploration and exploitation of marine resources,marine safety maintenance and other aspects is increasing.As a commonly used sensor for underwater detection,forward-looking sonar is not affected by underwater environment conditions.Compared with visible light and infrared light,it has a longer propagation distance and stronger penetration ability under water.But at the same time,sonar image has the characteristics of poor contrast,low resolution and serious noise,which makes the target detection and tracking in sonar image difficult to achieve and slow progress.Compared with traditional methods,deep learning method is more suitable for target detection and tracking of forward-looking sonar.Therefore,in this paper,a forward-looking sonar target detection and tracking method based on depth learning is proposed.The details are as follows:Firstly,the preprocessing method of forward-looking sonar image is studied.According to the storage structure of the forward-looking sonar data,the forward-looking sonar image is reconstructed,the bilinear interpolation is used to interpolate the forward-looking sonar image,and the segmented nonlinear gray-scale transformation algorithm is designed,and the pseudo color transformation is carried out for the forward-looking sonar image.Through experiments,the reliability of the selected algorithm is verified.Secondly,the target detection algorithm for forward-looking sonar is designed.By designing feature extraction network based on dense connection,adding feature pyramid network and designing feature mapping structure,a narrow channel and large depth target detection algorithm suitable for the characteristics of forward-looking sonar image is constructed,and a comparative experiment is carried out to verify the advantages of the algorithm in detection accuracy and speed.Thirdly,the target tracking algorithm based on twin network is selected to track the forward-looking sonar target.The main twin network target tracking algorithm is studied from three aspects: Twin network structure,target prediction structure and hierarchical aggregation mechanism.The twin network target tracking algorithm is studied from two aspects: the algorithm of selecting the region to be detected and the network structure of twin feature extraction The algorithm is improved.Finally,combined with the characteristics of the forward-looking sonar target detection and tracking algorithm,the fusion method of the target frame of the two algorithms is proposed.Then,aiming at the forward-looking sonar frogman target and the forward-looking sonar underwater small target,two commonly used target detection and tracking schemes and the scheme in this paper are set up for comparative experiments,which verify the feasibility of the proposed method.