Real time sensor fusion for autonomous underwater imaging in 3D

In this paper a method that enables real-time visual and acoustical processing for 3D imaging is presented. This, in turn, can be used for 3D-mapping of objects and environments, inspection and/or tracking or a ''lock and hold'' function for an Autonomous Underwater Vehicle (AUV). The described method gives a framework for real-time 3D information processing and modelling. It makes monitoring of 3D features easy and speeds up higher level algorithms. In underwater vision there is no single sensor that can give solutions to all perceptual problems that may arise. As in many other fields there is a need for data from several sensors to f.x. obtain the desired accuracy and range. The sonar has good range properties but may often have poor angular resolution. On the other hand the camera has poor depth-measurements capabilities and a rather limited sensor-range, but it has good resolution and is able to detect non-structural properties of the observed object such as colour, paint etc. An important property of this method is the fact that the raw-data from the sensors are processed on the fly. The relevant information extracted from the raw-data are organised in a 3D structure. This gives a better signal to noise ratio input to subsequent algorithms. The idea has been investigated with success for the 3D sonar case. Traditionally, the use of several sensors has meant ''use sensor B when sensor A fails''. Sensors have in other words been used purely complementary. In this method there is a close integration which is applied before the modelling, matching or identification of objects, making real-time 3D processing feasible at a low hardware cost. All in all this method has very good region-mapping properties in real-time and can be used in several possible areas. Methods like this is a necessary background for real-time 3D AUV perception.