Research On Key Technology Of Autonomous Navigation Algorithm Of Underwater Vehicle

With the increasingly fierce competition in the area of ocean development,demand in the various countries for the underwater robot which has higher capabilityof autonomous navigation is surging. In order to meet this, using data fusion methodinstead of equipping single sensor such as sonar for the robot has became the bestchoice. Simultaneous localization and Map Building is the key to implementautonomous navigation. The paper contains an in-depth study of the SLAM algorithmbased on filter and graph, an analysis of currently existing methods difficulties, aswell as improvement and innovation for this.First, the paper provides an overview of the robot s development, theautonomous navigation and the SLAM algorithms. Then by establishing thecoordinates, the robot-motion and the observation model, analyzes the weak points ofthe SLAM algorithm based on extended Karman Filter (EKF). After that, the datafusion pre-processing method on the basis of U-Karman Filter is proposed. Then thepaper introduces the SLAM framework on the basis of graph opt, including frame-to-frame alignment, loop closure detection and graph optimization. Finally by thestochastic gradient descent (SGD) algorithm and variable learning rate, it gives theonline graph optimization method.First, by establishing the robot-motion and the observation model, the paperintroduce the SLAM algorithms on the basis of EKF and then by analyzing its weakpoints on the aspect of computational efficiency and robot model linearization, itpropose the data fusion pre-processing method on the basis of U-Karman Filter, bywhich the covariance and the other Jacobian matrix s calculation during the updatephase will be reduced. Through two data fusion preprocess, including datapreprocessing and obstacle preprocessing, to reduce the number of operations, andthrough experiment simulation on off-line data to reduce the calculation and toachieve the effects of simulation results, the situation of repeatedly encountering thesame obstacle should be avoided.The paper showed that in the area of autonomous navigation the SLAMalgorithm based on graph opt has advantage over the SLAM algorithm based on filter. The conclusion is drawn by comparing their model processing, solve efficiency, andloop closure. Before establishing the robot pose graph to describe their nonlinearconstraint relations it needs to introduce the frontend frame of the SLAM algorithmbased on graph, including frame-to-frame alignment and loop closure detection, andalso needs to discuss the backend frame (graph opt part), including the modeling andthe optimization method. Finally by the stochastic gradient descent (SGD) algorithmand variable learning rate, it gives the online graph optimization method.The idea is: To give different learning rates to different parts of robot pose graphso that new constrain information can be disposed in time. By this way, when thelearning rate changes due to the next three reasons: first, the new emerging constrainmade the learning rate of the pose graph increase; second, exiting constrainprocessing made the previous situation of learning rate increase impact on the others;third, the learning rate is decreased by using the consistent harmonic series, it canmake the unaffected parts keep the low-error structural and also can make the partsbeing updated quickly find a new equilibrium position. And finally the effect of themethod, which is based on the various learning rate, is validate in the respect of thegraph s nonlinear constraint online solving by the experiment simulation results.