Research On Target Tracking Technique Of Forward-looking Sonar

Forward-looking sonar visual perception is an important branch of computer vision and artificial intelligence. It plays an irreplaceable role in the underwater robot system and it mainly undertakes the target detection, tracking, obstacle-avoidance and so on. Research on reliable underwater target tracking technique is of great significance no matter in military area or in civilian area. This dissertation aims at researching on basic ideas and fundamental methods of target tracking based on single forward-looking sonar. The image sequences are from two kinds of forward-looking sonar, DIDSON and Super SeaKing DST.Firstly, particle filtering principle is introduced to provide theory foundation for correlation tracking based on particle filter. On the one hand, Monte Carlo and Bayes Filtering, as the origin of particle filtering, is introduced. On the other hand, the degeneration problem is analyzed and solved.Secondly, some image preprocessing methods are researched according to forward-looking sonar imaging character. They are grayscale local enhancement algorithm, Otsu threshold segmentation, morphological processing methods and edge extraction. They can provide a reliable guarantee for target feature extraction and ameliorate the vision effect at the same time.Thirdly, the feature extraction on geometry, grayscale and statistical are researched respectively. The fusion of multiple features is the element of observation model.Finally, target tracking techniques include the correlation tracking and the correlation tracking based on particle filtering. This dissertation proposes tracking strategies respectively and focuses on robustness and tracking accuracy to evaluate their tracking performance. The correlation tracking deals with DIDSON sonar image sequences using template matching and its tracking effect is evaluated. The correlation tracking based on particle filtering deals with single target and double target from DST sonar image sequences. The system state transfer model, observation model and re-sampling model are established respectively. The character of particle filtering which has strong robustness and good performance on multi-peak processing is also analyzed. The final part consists of the particle radius adaptive adjustment strategy and template adaptive adjustment strategy which can further enhance the robustness and the tracking accuracy.