Perception methods for holedetection in mining

Sammanfattning

In this work we attempt to offer improved hole detection approaches for underground robotics using the pre-existing object detection methods, (You Only Look Once, version4) YOLOv4 and a contour detector. In the process, we created a training set and test set using a (Light Detection And Ranging) LiDAR device on a testing wall with holes on itsface, and trained YOLOv4 models on Red Green Blue (RGB), gray-scale and depth images resulting in three YOLOv4 models for each category. We then combined the outputs ofthe YOLOv4 on RGB with the output of YOLOv4 on depth images as well as distance values measured by the LiDAR, resulting in a combination of Red, Green, Blue and Depth (RGBD). The RGB data was used as benchmark. That is, all other approaches’ outputs were compared with it. As a result, three novel combination algorithms were successfully created: Distance Combo, Overlap Combo and Contrast Combo. With a true positiverate at 75.26%, false negative rate at 24.74% and false positive rate at 0,96% Distance Combo outperformed all other approaches including the SimpleBlobDetector which was used as an ellipse detector.