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

At present,the forward-looking sonar is mainly responsible for target detection,target tracking and other tasks in the water,and it plays an important role in the underwater submersible system.However,due to the low resolution of the forward-looking sonar images and the serious noise interference,the non-rigid deformation of the target changes the apparent model of the target,and the small targets similar to the background are extremely difficult to distinguish.These interference factors make the front-looking sonar underwater target tracking tasks have become very difficult.In this paper,different depth learning methods are used to track the larger frogman targets that are likely to be deformed in the forward-looking sonar and the small targets that are easily confused with the background.First of all,we briefly introduce the development process of deep learning,and then the current research status of target tracking technology at home and abroad and the current problems of forward-looking sonar target tracking technology are elaborated.And we use the adaptive threshold segmentation method and the Kalman filter tracking method to point out that the traditional target tracking method can not overcome the noise interference well,and there is a problem of false detection,resulting in tracking failure.Therefore,this paper proposes a deep learning method to solve the problem of visual target tracking.Secondly,create an image dataset.The high cost of the forward-looking sonar image acquisition results in a small amount of data available,and the image after imaging is severely affected by noise,resulting in extremely poor image quality.For these problems,an image processing method using geometric transformation and color adjustment is proposed to expand the data set.The high-quality data is also guaranteed while the implementation of the deep learning method is guaranteed.Thirdly,the study of lightweight convolutional neural network target tracking methods.Then,the non-rigid changes of moving large targets in the forward-looking sonar image often change the apparent model of the target.The complex network structure will reduce the real-time performance of the tracking system.Therefore,this paper proposes a lightweight-based convolutional neural network model to solve these problems.Finally,the study of the fully-convolutional Siamese network target tracking methods.This paper analyzes the full convolutional neural network and the siamese network,and predicts the location of the target based on the probability of a certain pixel location to predict the location of the target for tracking.A target tracking model based on a fully-convolutional siamese network structure is used to solve the tracking problem of small targets in the forward-looking sonar images.In this paper,through multiple sets of contrast experiments,it is proved that the improvement of the network structure for the specific targets appearing in the forward-looking sonar can make the deep learning method more flexible.At the same time,the feasibility and effectiveness of the proposed target tracking method based on lightweight convolutional neural network and deep learning target tracking method based on full convolution siamese network are proposed to improve the accuracy and efficiency of target tracking.Finally,the paper points out the existing problems and briefly explains the future development direction of deep learning and target tracking.