Underwater SLAM for Structured Environments Using an Imaging Sonar

More than 70% of the earth’s surface is covered by water. Oceans and seas host an incredibly rich biodiversity, influence short and long term climate and have a high impact not only on the economy, but on the life and evolution of human society as a whole. Exploring this large body of water is a matter of the utmost importance, not only because it represents a vast source of natural resources, but also because its study may help us understand how this complex ecosystem works.

Simultaneous Localization and Mapping (SLAM), also referred to as Concurrent Mapping and Localization (CML), is a fundamental problem in mobile robotics that has been the focus of substantial amount of research work in recent years [30, 2]. The objective of SLAM is to make it possible for a moving robot starting at an unknown location without previous knowledge of the environment to build a map using its onboard sensors while, at the same time, using this same map to compute the robot’s location. Although performing these two tasks simultaneously may seem complex, the essentials behind SLAM are indeed quite simple. Figure 2.1 illustrates the basics of the process. A moving vehicle will inherently accumulate errors in its position estimate as a consequence of the noise introduced in the deadreckoning and/or the inaccuracies in the use of a prediction model. Moreover, errors will also affect the map building process. The sensor that perceives the environment is mounted in the vehicle and therefore its position uncertainty will be incorporated when new information is added to the map. As a result, the vehicle will eventually get lost and the map will become unusable (Figure 2.1(a)). A system performing SLAM, however, is able to attenuate and even contain this uncertainty growth by means of the reiterated observation of the elements stored in the map. Figure 2.1(b) represents the situation where a new measurement from the robot is likely to correspond to an entity already incorporated in themap. Then, a data association process should be carried out to determine the correct matching. When this process is positive, this information is used to update the estimates of both the vehicle’s position and the map. Adding more information results in a better estimate and hence a reduction of the uncertainty in the problem (Figure 2.1(c)).

This chapter describes the Ictineu AUV (Figure 3.1), the research vehicle of the Computer Vision and Robotics Research group at the University of Girona that constitutes the experimental platform of this thesis.

In 2006, the Defence Science and Technology Lab (DSTL), the Heriot-Watt University and the National Oceanographic Centre of Southampton organized the first Student Autonomous Underwater Challenge - Europe (SAUC-E) [27], Europeanwide competition for students to foster research and development in underwater technology. Ictineu AUV was originally conceived as an entry for the SAUC-E competition by a team of students collaborating with the Underwater Robotics Laboratory [129, 109]. This author, who was team leader during the competition, became involved in the hardware design and construction phase as well as in the development of the sonar based localization system (described in Section 5.2). Although the competition determined many of the vehicle’s specifications, Ictineu was also designed taking into account its posterior use as an experimental platform for various research projects in our laboratory. The experience gained by the group in the previous development of vehicles such as the Garbi AUV, made it possible to build a low-cost vehicle of reduced weight (52 Kg) and dimensions (74 x 46.5 x 52.4 cm) with remarkable sensorial capabilities and easy maintenance.

The purpose of this chapter is to give a brief introduction to the operational principles of MSISs by explaining the basics behind the acquisition of acoustic images as well as providing tools to understand and interpret the information they contain. Moreover, some hints about the principal issues associated with managing MSIS data are given at the end of the Chapter. Some of the figures and examples described here are adapted from the introductory document in [59]. A deeper study on sonars and their techniques can be found in [128].

Section 5.1 reviews different strategies to perform data association in localization problems, while Sections 5.2 and 5.3 present two map-based localization methods developed for the competition. The first is a simple algorithm which determines the vehicle’s position by means of a voting strategy, while the second relies on an EKF to merge the information from several sensors and the tank map. A third method, which combines several aspects of the two other algorithms to improve the estimation process, is presented in Section 5.4. The chapter concludes with a summary of the advantages of the different methods and some guidelines for further work.

In this chapter a SLAM framework for AUVs equipped with an MSIS operating in manmade structured environments is proposed. In the previous chapter, the use of techniques such as the Hough transform and the Kalman filter were studied in the context of a localization problem. Here, these techniques are further explored for their application in SLAM. The proposed approach is composed of two parts running simultaneously. The first is a line feature extraction algorithm which is responsible for managing both the measurements arriving from the MSIS and the vehicle position estimates from the SLAM system to search continuously for new features by means of a voting scheme. Eventually, when a new feature is detected, the algorithm also estimates its uncertainty parameters through an analysis of the imprint left in the acoustic images. The second part is a Kalman filter implementation which is the core of the proposed SLAM system. This filter merges the information from various sensors (DVL, compass and pressure sensor) and the observations from the feature extraction algorithm in order to estimate the vehicle’s motion and to build and maintain a feature based map (see Figure 6.1 for a diagram of the complete system). In addition, the problems associated with large scenarios have also been addressed through the implementation of a local map building procedure. At the end of the chapter, two tests performed with real sensor data endorse the proposed SLAM approach. The first employs the dataset corresponding to the previously presented CIRS water tank test, while the second undertakes amore realistic application scenario with a dataset obtained in an abandoned marina.

This concluding chapter summarizes the thesis by reviewing the contents described in each chapter. The significant research contributions are then listed. The objectives still to be accomplished and interesting future research issues are discussed in the future work section. The research framework for the thesis is then described. Finally, the publications related to this work are listed.