• Document: 6 DOF EKF SLAM in Underwater Environments
  • Size: 5.4 MB
  • Uploaded: 2019-04-16 08:06:25
  • Status: Successfully converted

Some snippets from your converted document:

6 DOF EKF SLAM in Underwater Environments MARKUS SOLBACH Universitat de les Illes Balears Abstract. The increasing number of industrial or scientific applications of 1. INTRODUCTION Autonomous Underwater Vehicles (AUV) raises the challenging question on how to derive the vehicle’s localization accurate enough for the mission 1.1 Problem Statement success. In the last years, technological advances made easier the accessibil- This paper details an approach to accurate localization based on EKF (Ex- ity of the sub-aquatic world for research, exploration and industry tended Kalman Filtering) SLAM (Simultanously Localization and Map- exploitation. Nowadays Remotely Operated Vehicles (ROVs) are ping) with pure 3D stereo data, which consists of three major stages. used in a wide range of applications, such as maintenance, res- Stage one is, in terms of EKF, the so called prediction stage. During cue operations, surveying, infrastructure inspection and sampling. this stage the algorithm predicts the vehicle’s localization using the visual Some of ROVs limitations, such as limited operative range and the odometry, which is known to be noisy and to provide drift in position and need of support vessels, are overcome by Autonomous Underwa- orientation (pose). The uncertainty of the odometry data is modeled with ter Vehicles (AUVs). These kinds of vehicles are used in highly the covariance matrix. repetitive, long or hazardous missions. Moreover, since they are Stage two is the state augmentation step. In this phase, the current odometry untethered and self-powered, they are also significantly indepen- estimation is added at the end of the state vector of the EKF. The uncertainty dent from support ships and weather conditions. This, in compar- accumulated over time makes the resulting predicted state non reliable. ison to ROVs, can reduce considerably the missions costs, human The last Stage (update) tries to reduce this error by finding visual Loop resources and execution time. Closings. Loop Closings are areas of the environment which the robot al- One of the most challenging points in research associated to un- ready observed in the past. Loop Closings are important because they pro- derwater vehicles is the one of localization. There are several pos- vide the system with new and often more reliable information, what is a sibilities to estimate the vehicle’s pose, for instance, using inertial second transformation of an already observed one. With the difference of sensors, or by computing the odometry with acoustic sensors or these two transformations the approach is able to update the whole state cameras. Another possibility is sensor fusion, which means com- vector to a one with less error by using Extended Kalman Filtering equa- bining inertial sensors and odometers, in extended Kalman filtering tions. (EKF) or particle filters, to correct errors within the trajectory [Lee During the three steps of the filter, all the data concerning the robot pose et al. 2004], [Kinsey et al. 2006]. (odometry and filter estimation) are expressed as (x, y, z) for translation The most successful approach to perform a reliable localization is and a quaternion (qw , q1 , q2 , q3 ) for orientation. called SLAM [Durrant-Whyte and Bailey 2006]. SLAM (Simulta- In this work, a pure stereo system is used to compute the visual odometry neous Localization And Mapping) computes the position of the ve- in 3D and to find the visual loop closings. A Kalman update is performed hicle simultaneously to the calculation and refinement of the posi- if the algorithm is able to find a Loop Closing between an image associated tion of landmarks of the environment. to a position stored in the state vector with the current image, that is, if both In the past, underwater SLAM was mainly developed by using images present a certain level of overlapping. acoustic sensors. These sensors provide good underwater proper- To calculate the motion of the camera between two positions which are sus- ties, such as large sensing ranges[Ribas et al. 2007]. The problem pected to be a loop closing, first SIFT [Lowe 2004] features or SURF [Bay with acoustic sensors is the spatial and temporal resolution, which et al. 2008] features of both images are computed. Then, applying the prin- is lower than using cameras. This higher resolution permits cam- ciple of the stereoscopy, the 3D points corresponding to the image features eras to provide more environmental data than the data provided by matched between the current stereo pair are cal

Recently converted files (publicly available):