Study On Underwater Target Detection And Tracking Of Autonomous Underwater Vehicle

With the ocean exploitation developing, more and more autonomous underwater vehiclesare manufactured. As one of the most important sensors of AUV to perceive the externalenvironment, forward-looking sonar is used for avoiding obstacle, navigation, path planning,environmental detection, etc. However, the acoustic image’s rougnness results in that a lot ofgeneral image processing and target tracking technology can not achieve the purposeeffectively, it is necessary to learn and research the problem.The forward-looking sonar’s imaging characteristics and image processing technologyused in sonar image segmentation are studied, and their characteristics are analysed andcompared. The combination of the entropy concept and threshold segmentation theory isstudied to build the1D maximum entropy and2D maximum entropy threshold method.Combined with fuzzy theory,2D fuzzy entropy threshold segmentation algorithm isintroduced, and the algorithm is improved in the sonar image segmentation experiments. Todetect the moving targets in sonar image, the traditional inter-frame difference method andentropy threshold method are combined to establish the spatial-temporal histogram,2Dspatial-temporal entropy threshold algorithm and2D spatial-temporal fuzzy entropy thresholdalgorithm are proposed respectively. Then they are used to detect moving object inforward-looking sonar image sequences. Finally, the improved method based on2D fuzzyentropy threshold algorithm, region growing method and2D spatial-temporal fuzzy entropythreshold algorithm are compared, the results show that the last method can detect the movingobjects better; furthermore it suppresses background noise effectively.Combining image analysis technology with forward-looking sonar’s imagingcharacteristics, the expression of the underwater target features in the image sequence isstudied, and various features of the three consecutive frames are calculated. The changes ofthe typical targete feature in different sample images are compared and analysed. The mostbeneficial feature combination for detection and tracking is determined using generalizedregression neural network.Kalman filter and its extended algorithms are studied and applied to underwater movingtarget. Then aimed at the nonlinear estimation problem, Bayesian estimation and particle filter theory are researched and analysed. The underwater moving target motion model isestablished, and a single target motion tracking is achieved based on the selected features andinformation fusion strategy. The representative tracking figures, and the comparison of actualtrajectory and filtering results of the target are drew. The influence on PF results of severalparameters is compared and analyzed. The comparison shows the superiority of PF fornonlinear system estimation problem with EKF filter tracking results.