Evaluation of vision-based surface crack detection methods for underground mine tunnel images

Safety in underground mines is an important aspect for mining companies. Geological failures such as roof collapse and rockfalls are one of the most fatal safety hazards in an underground environment. A way to prevent such hazards is to detect early signs of geological failures and hence implement safety measures. The presence of surface cracks is one of the early signs of a geological failure and the current method of detecting them is by sending a geotechnical engineer to survey underground tunnels. This is a risky operation due to the unpredictable hazards and harsh underground environment. Deploying a remote vehicle attached with suitable sensors with the ability to autonomously detect early signs, such will mitigate such risk and also assist geologists to interpret massive amount of data quickly. Several vision-based methods to automatically detect cracks in images can be found in the literature, however, no indication of the performance of such methods in the context of underground mines is available. This paper provides an experimental evaluation of those methods on images collected in a real underground mine. The results show that existing methods perform relatively poorly in this context, indicated by an F1 score ranging between 20% and 63%.