Research On Underwater Single Target Tracking Algorithm And Experiment Based On Siamese Convolutional Neural Network

The ocean is rich in biological and mineral resources.Using deep-sea equipment to track underwater organisms and the monitoring and tracking of deep-sea mining processes is of great significance to human exploration of marine resources.How to track the underwater single target objects quickly,accurately and stably has become popular topics studied for scholars in recent years.There are certain particularities and difficulties in underwater single target tracking,such as poor clarity and contrast of underwater images,complex and changeable background environment,the movement of the object itself and being blocked,etc.In recent years,machine vision and deep learning have developed rapidly.The single-target tracking algorithm with deep learning as the core is significantly better than the traditional single-target tracking algorithm when processing natural images in complex environments.Therefore,aiming at the problem that the complex underwater environment seriously affects the accuracy,stability and real-time performance of underwater single target tracking,this paper is based on the siamese convolutional neural network SiamRPN++algorithm for underwater single target tracking research.The main contents are as follows:(1)Aiming at the problem that the underwater image has the difficulty of extracting features due to interference factors such as poor clarity and contrast,complex and changeable background,this paper increased the number of Conv4_x layers from 50 to 62 and designed SiamRPN++based on ResNet-62 backbone network structure.Using VOT2016's five groups of underwater single target tracking videos as test video sets,and quantitatively analyzing the algorithm.The results showed that compared to the original algorithm,SiamRPN++with ResNet-62 as the backbone network has improved accuracy by 20.53%,robustness by 9.07%,and EAO by 12.99%.(2)Aiming at the problem that there are too many parameters and too much calculation in the original algorithm and the hardware limitations of underwater equipment,this paper learned from the designing idea of the inverted residual bottleneck block in the modern network MobileNetV2 and proposed a new type of NewNet-50 backbone network structure,and it was quantitatively analyzed using the evaluation indexes of(1).The results showed that compared to the original algorithm,SiamRPN++with NewNet-50 as the backbone network has improved accuracy by 12.63%,robustness by 9.07%,EAO by 53.11%and reduced network complexity by 49.34%,network parameters by 37.85%.(3)In order to achieve the purpose of both improving the backbone network's ability to extract underwater image features and reducing network parameters and calculations,that is to simultaneously improve the accuracy,robustness and real-time performance of the algorithm,combining the algorithms of(1)and(2),using the network layer distribution of(1)and the basic building blocks of(2)to design the NewNet-62 backbone network structure,and it was quantitatively analyzed using the evaluation indexes of(1).The results showed that compared to the original algorithm,SiamRPN++with NewNet-62 as the backbone network has improved accuracy by 37.89%,robustness by 18.15%,EAO by 71.19%and reduced network complexity by 37.95%,network parameters by 22.70%.(4)This paper did an experimental study of SiamRPN++tracking algorithm with NewNet-62 as the backbone network.Testing the tracking effect by setting up underwater experiment hardware and software platforms,and setting the tracking object scale change,self-rotation,tracking object being blocked,tracking object deviating from the field of view,reflection and other interference factors in the experiment process.The results showed that the algorithm has good tracking performance.