Deep Learning And Research Of Robotic Fish Target Recognition Based On Cloud Computing Platform

This dissertation takes the bionic robotic fish as the research object,and studies the distributed parallel deep learning recognition method of robotic fish.The target recognition and tracking control of robotic fish is mainly to extract the robotic fish from the complex environmental background,and send out control signals to the robotic fish according to the surrounding environment and its speed,posture and other relevant information.Due to the complexity of the background,changes in ambient light,target movement,and many other factors,there are many problems in target recognition.In current robotic fish target recognition algorithms,algorithms based on color information and interest point detection are commonly used recognition algorithms.But experiments have shown that methods based on color and contrast are susceptible to illumination and irregular reflection of water surface.Based on the existing experimental environment and conditions,a global visual camera and deep learning technology are adopted to carry out research on target recognition and control of multi-bionic robotic fish.The main research results can be summarized as follows:Firstly,based on the existing global visual robotic fish control platform,through a secondary development,a stable acquisition of moving images of robotic fish is obtained.The machine fish coordinates are acquired according to the template matching method to identify and track the positioning robotic fish.And the machine is acquired according to the serial port.The fish’s direction gear and speed gear are used as identification tags to prepare data for deep learning and recognition of robotic fish.Secondly,the deep-learning method is used to perform convolutional neural network training on the moving images of robotic fish so that it can reach the training target for basic recognition of robot fish poses.Thirdly,multi-machine distributed parallel computational storage experiments were conducted.A distributed computing cluster composed of multiple lower hardware configuration machines was used to improve the timeliness and accuracy of singlemachine fish image recognition.The distributed parallel scheme is implemented in the form of HDFS and the remote call mode of RPC combined with Thift as well as the structural characteristics of the convolutional neural network are suitable for parallel training of the model,and the multi-machine parallelism and network parallelism are realized.It improves the utilization of computing resources in the laboratory and optimizes the robot fish identification model.And providing basic conditions for subsequent online training to identify robotic fish models.Fourth,the distributed parallel recognition module is rebuilt to replace the image recognition and template matching location tracking part of the original global visual robot fish control platform.The speedup of program optimization and recognition efficiency is obvious.Fifth,the design of a new distributed parallel online training and recognition control platform system experiments.The system experiment consists of a global visual recognition control server and a distributed parallel recognition module.The global visual recognition control server is stripped from the previous global visual robot fish control platform.The distributed parallel recognition module is provided by a multi-machine distributed parallel computing storage section.Sixth,carry out deep-learning recognition control experiments for robotic fish,water polo,and multi-machine fish control to obtain the desired experimental results.After training,the recognition network has strong feature abstraction capabilities,and can continuously estimate the position of the tracking and positioning locomotive fish.And can adapt to different light changes in the environment and other moving objects interference.Finally,the relevant research work completed in this thesis is summarized and analyzed.And the next research direction and application prospect of the global visual bionic machine parallel recognition control system based on cloud computing are pointed out.