Research On Sonar Image Target Recognition And Tracking Based On Deep Learning

The ocean is the Treasury of mankind,containing huge energy and resources.Human activities under water are limited,so intelligent robots and detection equipment should be developed to replace human underwater operations.The underwater unmanned vehicle is an excellent underwater mobile observation platform.It has a unique movement form and can carry sonar for underwater observation to overcome the drawbacks of traditional observation methods.This thesis mainly studies the underwater sonar image preprocessing methods,target recognition and tracking methods.The main research contents are as follows:Firstly,the sonar image segmentation method is studied.Several classical underwater image segmentation and recognition algorithms are summarized and verified,and the practical disadvantages are compared.Considering that sonar background noise is large in the shallow water area of the ocean,and is easily affected by seabed reverberation to reduce the signal to noise ratio,this thesis proposes an improved Otsu algorithm,which can effectively enhance the image contrast and accurately segment the target area.Secondly,the method of sonar image target recognition is studied.On the sonar image by the superposition of a variety of noise interference,and exist in the noise reduction result target image blurring is aggravating,and even a large number of image signal loss problem,puts forward the improved adaptive threshold denoising NSCT transform as sonar image preprocessing,and focused on the highlighting of sonar target image and the characteristics of the small target,An end-to-end improved SSD sonar image underwater multi-scale target detection model is proposed.The basic network is further extracted by multi-branch and multi-scale method,and the weights of target features are shared.Finally,the algorithm is evaluated from two indexes of average accuracy value and time.Thirdly,the sonar image target tracking method is designed.Based on image information data and deep learning model,design the particle filtering fusion extended Kalman filtering method,forecast the state of the dynamic target obstacles,the filtering method based on improved SSD model and design,combined with the characteristics of target recognition and tracking algorithm,through the observation and prediction of obstacles of fusion algorithm of target state estimation calculation,In order to reduce the false recognition rate of the tracked object,the experiment proves that the tracking algorithm has a good effect.Finally,an experimental demonstration system for sonar target recognition and tracking is developed.Based on the original data collected by sonar,the algorithm proposed in this thesis is verified.The data set test shows that the above algorithm can meet the real-time and accuracy requirements of sonar image target recognition and tracking system,which indicates that the algorithm proposed in this thesis is feasible and effective.