Visual odometry and image mosaicing for NDE

Miniature robotic vehicles are receiving increasing attention for use in non-destructive testing (NDT) due to their attractiveness in terms of cost, safety and their accessibility to areas where manual inspection is not practical. Cameras can play several roles in a robotic system as a result of the information rich nature of the 2D image signal including object/shape recognition, motion estimation and ranging. This paper considers a robot configured with a single camera that monitors the surface of the structure under the robot as it moves. Visual odometry is demonstrated that can reliably track the robot's position. An image mosaicing approach is presented that calculates the robot's path using the complete image set, taking account of overlapping images that were not consecutive in the image stream. It is shown that this significantly reduced the absolute error of the positional estimate. The system was found to work well on surfaces that were not highly reflecting including steel which is common in industrial plant.