Research On Underwater Container Identification And Three-dimensional Reconstruction Based On Multi-beam Water Column Data

Shipping has the advantages of low cost and large transportation capacity,which is of great significance to the economic and social development of the basin.With the increase of the number of ships,the danger of container falling into water caused by external factors or human factors occurs from time to time.After a dangerous situation occurs,quickly locating the falling container and salvaging it in time will help to ensure shipping safety and reduce the risk of environmental pollution of water.At present,the main means to locate the falling container is underwater terrain survey,which depends on manual interpretation.It has low precision,low efficiency and limited automation,and can not effectively reproduce the three-dimensional shape of the target in the water.The water column data collected by the multi beam system has three-dimensional imaging ability and can accurately detect the falling container.Therefore,aiming at the multi beam water column data,this paper identifies the falling container,and studies the target extraction and three-dimensional reconstruction.The main contents include:(1)Multi beam water column data processing.This paper introduces the composition of multi beam sonar system,and expounds its imaging mechanism;Taking the S7 K format data collected by Sea Bat series multi beam sonar system as an example,the water column data is processed by data decoding,position reduction,image interpolation and image denoising to obtain a clear water image,which lays a foundation for the subsequent multi beam water image falling container detection.(2)Underwater container target recognition based on improved UNET network.In order to realize the automatic detection of falling container in water image,the classical UNET network model is improved.The model takes the pre trained vgg16 network as the backbone feature extraction network of the classical UNET network model;Focal＿Loss,as the loss function of the model,balances the proportion of each segmentation target in the image;Model training combined with transfer learning strategy can improve the training effect of small sample training data set.Through experiments,the effects of the improved UNET,the classical UNET,deeplabv3 +,and the improved UNET model with transfer learning strategy on the segmentation of falling container targets are compared and analyzed.The results show that the average accuracy(MPA)of the above four methods are61.92%,90.80%,91.06% and 92.38% respectively.Compared with the classical UNET network,the MPA accuracy index of the improved UNET model is improved by 47.1%,the class average intersection union ratio(Miou)index is improved by 82.3%,and the overall segmentation performance is greatly improved.The migration learning strategy will increase the MPa and Miou indexes of the improved UNET network by more than 1%.(3)Research on 3D reconstruction technology of submerged container target.A three-dimensional reconstruction method of falling container target based on the shortest diagonal sequence contour is proposed.Considering the characteristics of multi beam water column data Ping by Ping acquisition,the sampling points of falling container are Ping stitched one by one to reconstruct the target.By extracting the peripheral contour points of the image with improved UNET network semantic segmentation,and after manual editing and coordinate conversion,it is used as the sampling point of target 3D reconstruction.The experiment compares and analyzes the three-dimensional reconstruction method of sequence contour with Power Crust,the three-dimensional reconstruction method of poisson surface.It is found that this method has high efficiency,the time required is only 1 / 11 of Power Crust algorithm,and the three-dimensional reconstruction effect is good,which can effectively reproduce the position and sink shape of the submerged container.(4)Design and implement multi beam water column data processing software.The MFC platform of Visual Studio designs and develops a set of falling container detection software.The software includes the functions of image imaging of water column data,water column data analysis and falling container detection and extraction.Using the measured multi beam water column data,the function of the software is introduced.