Research On Deep Learning Target Tracking Methods For Underwater Robot In Complex Environments

Underwater target tracking is an indispensable technology for underwater robot in operation,which provides data support for its subsequent path planning and operation control.The natural underwater environment is more complicated,there is occlusion caused by underwater objects and the influence of ocean currents.They are the challenges for the target tracking method.At present,the tracking methods used in underwater environment are mostly based on correlation filtering methods,but these traditional underwater target tracking methods have problems in accuracy when facing complex environment,and their learning ability is limited.With the advancement and development of deep learning,it has gradually become a popular research direction in the field of target tracking.Then in the complex underwater environment,the target tracking methods based on deep learning will also become a future research trend.In this paper,the tracking method based on deep learning framework is applied to the underwater complex environment.The problems and the effect of the modified algorithm are analyzed and studied.The main contents of this paper are as follows.Firstly,an underwater target tracking algorithm based on improved siamese network is proposed.Based on the convolution neural network as the feature extraction network,a batch normalization layer is added to reduce the problem of gradient explosion or gradient disappearance,and a pyramid pool layer is introduced to deal with the problem that the network parameters need to be modified many times due to different image sizes during underwater data acquisition,so that the underwater images and their features can be fully utilized,In addition,an adaptive training strategy is proposed to reduce the negative impact of underwater tracking.Compared with the traditional algorithm,this algorithm improves the success rate and robustness of the tracking algorithm to a certain extent.Secondly,referring to the structure of more advanced deep learning target tracking method,an underwater target tracking method based on lightweight neural network is proposed.The deeper feature extraction network framework is introduced into the underwater tracking method.In order to deal with the problem of over fitting caused by the contradiction between the deepening of the network and the few features of underwater targets,the lightweight neural network is also introduced to reduce the amount of calculation and parameters of the network,and the lightweight neural network module is used to replace the convolution layer in the original network.In addition,in order to adapt to the characteristics of underwater vehicle working for a long time,a stable tracking mechanism is designed.Using the motion characteristics of underwater vehicle and the characteristics of target tracking task,the tracking range is reduced to ensure the continuity and stability of tracking task.The amount of computation and model parameters of the algorithm are alleviated,and the tracking effect is significantly improved compared with the traditional underwater target tracking methods.Finally,in order to further improve the performance of underwater target tracking algorithm,an underwater target tracking algorithm based on hybrid excitation model is proposed.In this method,the hybrid excitation mechanism is introduced to extract and fuse the features from the three aspects of space-time,channel and motion.The hybrid excitation model is placed in the front end of the bottleneck block in the residual network to improve the learning and understanding ability of the target.Aiming at the problem of water quality turbidity caused by underwater current,a tracking and prediction strategy is proposed to ensure the robustness of tracking task under the influence of current.The robustness of the algorithm is verified by experiments on underwater datasets and public datasets.