A Complementary Vision Strategy for Autonomous Robots in Underground Terrains using SRM and Entropy Models

This work investigates robots' perception in underground terrains (mines and tunnels) using statistical region merging (SRM) and the entropy models. A probabilistic approach based on the local entropy is employed. The entropy is measured within a fixed window on a stream of mine and tunnel frames to compute features used in the segmentation process, while SRM reconstructs the main structural components of an imagery by a simple but effective statistical analysis. An investigation is conducted on different regions of the mine, such as the shaft, stope and gallery, using publicly available mine frames, with a stream of locally captured mine images. Furthermore, an investigation is also conducted on a stream of dynamic underground tunnel image frames, using the XBOX Kinect 3D sensors. The Kinect sensors produce streams of red, green and blue (RGB) and depth images of 640 × 480 resolution at 30 frames per second. Integrating the depth information into drivability gives a strong cue to the analysis, which detects 3D results augmenting drivable and non-drivable regions in 2D. The results of the 2D and 3D experiment with different terrains, mines and tunnels, together with the qualitative and quantitative evaluations, reveal that a good drivable region can be detected in dynamic underground terrains