Object detection and classification in 3D sensor data in the surfacemining environment

Sammanfattning

This thesis investigates the integration of 3D sensor data with machine learning algorithms for object detection and classification in the surface mining environment, with the aim of supporting the transition to safer and more efficient autonomous mining operations. Conducted in collaboration with Epiroc, the project evaluates the feasibility, precision, and reliability of detection models under the harsh and dynamic conditions typical of mining sites, including dusty environments and varying terrain. Data were collected at Epiroc’s test site using an Ouster OS0-128 Light Detection and Ranging (LiDAR) and a ZED2i stereo camera. The datasets were annotated and used to train and evaluate three models: Point-Voxel Region-based Convolutional Neural Network (PV-RCNN) and Sparsely Embedded Convolutional Detection (SECOND) for point cloud data, and You Only Look Once version 11 (YOLOv11) for signal image data. The models were assessed on multiple metrics including mean Average Precision (mAP), Bounding Box (BBOX), Bird’s Eye View (BEV), and 3D Average Precision (3D AP). Results show that while high BBOX accuracy was achieved across all models, performance in BEV and 3D AP was notably lower, particularly under unfamiliar or dusty conditions. Among the tested models, PV-RCNN demonstrated superior general performance especially in terms of robustness across varied scenarios. While the image model performed well in specific scenarios. The outcomes confirm the potential of 3D LiDAR based object detection in autonomous surface drilling operations.