Research On Optical Positioning Method Of Underwater Nodes Based On Neural Network

Underwater optical communication plays an important role in the development of marine resources.In order to make underwater communication more efficient and stable,the transmitter of the information needs to estimate the receiver position in advance to establish an optical transmission link more accurately.That is,before the underwater optical communication,the transmitter first needs to locate the receiver.At present,the main underwater node positioning scheme mainly adopts a centralized deployment scheme,where multiple reference nodes are fixedly deployed and wired,and the centralized positioning is performed at the central node of the network.In the centralized scheme,the reference nodes need to be accurately located and the information is highly synchronized,which is difficult in the deployment and maintenance of the node.Therefore,two distributed underwater node positioning schemes where the underwater nodes can collect data and run the localization algorithm independently are propsed in this thesis.The main work and innovation of this thesis are mainly reflected in the following three aspects:Firstly,a laser localization scheme for underwater node based on RSS is proposed,where the node to be located emits a laser beam,and the node running the localization algorithm is equipped with a light intensity receiver.When the relative position of the two nodes changes,the received light intensity will also change accordingly.Therefore,in this thesis,the light intensity sequence values at relative different positions are used as the features of machine learning to train two localization network models and the results are evaluated.The experimental results show that when the relative distance between the two nodes is within 15 m,in the environment with water quality attenuation coefficient of0.151,the localization MSE of the model trained by the BP neural network can reach0.71 m.Secondly,an LED image-based underwater node localization scheme is proposed,which can be transferred to the underwater environment by collecting node image data in the atmosphere.In this thesis,six image detection models,SSD,Faster-RCNN,YOLOv3,YOLOv4,YOLOv5 and YOLOX,are trained respectively to identify target node in the collected images.Among them,YOLOX,the best performing model,has an AP of99.92% for node detection in the atmosphere,and an AP of 89.77% for underwater node detection.At the same time,a monocular vision localization algorithm based on similar triangles and graph fitting is proposed.The target node can be located by the camera focal length,the real size of the target and the pixel size of the target which is calculated by polynomial fitting of the LED strip curve.Finally,the test of the underwater node localization scheme based on LED images is carried out.The node image datasets in the atmosphere and underwater are respectively established,where the maximum distance between the camera and the node is 18 m,and the underwater node images are supplemented by Cycle-GAN.The MSE of the localization algorithm was tested to be 0.86 m.