Visual Localisation for Pipe Inspection Robots using Prior Information of the Environment

The ability of autonomous vehicles to explore and navigate their environments has improved greatly in recent years. However, many challenges remain to be overcome and many environments exist that present unique challenges to the problem of autonomous navigation. One such environment is subterranean pipe networks, such as sewers. There is a necessity to inspect these environments for maintenance purposes and, due to the difficult and time consuming nature of manual inspections, a desire to automate the process with mobile robots. As such, the challenges presented by subterranean pipe networks to autonomous navigation must be overcome.

This thesis contains a review of the existing research on SLAM algorithms with a focus on their use in pipe networks. The review aims to understand the state of the art approaches to solving the SLAM problem and how those approaches may be applied to pipe networks. The included review also highlights the specific properties of pipe networks that present additional challenges to SLAM algorithms that may not be encountered in more general environments. Of those presented in the review section, the principle challenges this thesis is concerned with stem from the relatively featureless and uniform pipe walls and the lack of communication with external systems. These characteristics make the feature detection and matching used by most visual SLAM algorithms difficult and the positional correction provided by systems such as GPS impossible.

This work attempts to overcome these challenges by exploiting the predictability of landmarks within pipe environments and the ability to know or create maps of those landmarks prior to navigation within the pipes themselves. This is presented in two publications, the first of which details a method of visually recognising joints and manholes and associating them with their known locations in a prior map to perform odometry. This approach has lead to a navigational system able to achieve a Mean Absolute Error up to 6 times lower than a state of the art algorithm in the same environment and which is capable of operating in environments where the state of the art algorithm fails entirely.

The second publication presents a system of further robustification in full networks by representing the results of an odometry system as a particle filter loosely constrained to a simple map. The system accounts for uncertainty in the path taken by the agent within the map by instantiating multiple particle filters representing different hypotheses and then assessing their probability of being the true trajectory based on the odometry outputs and the map. This allows the system to recover from failure and also improves the mean position and heading errors of the original odometry system by an order of magnitude and the final positional error by three orders of magnitude.

These publications represent a step toward the autonomous inspection of pipe networks and demonstrate the effectiveness of taking non standard and more specialised approaches over more commonly used and generalised methods when attempting autonomous navigation in uniquely challenging environments.