Comparison of 3D object detection methods for people detection in underground mine

Abstract

Autonomous vehicles have received immense attention in the mining industry nowadays. However, there is limited research on 3D object detection in the underground mine. This thesis wants to compare the ability of 3D object detection models in the underground mine environment. Three state-of-the-art 3D object detections are analyzed to detect people in the underground mine. The author collects 1000 point cloud files from Edgar mine to train and test the algorithm performance. Data labeling and preprocessing methods are discussed to convert raw point clouds to the algorithm's format. Then, several training parameters such as the number of datasets and epochs are analyzed to obtain the maximum performance of object detection methods. PV-RCNN has the highest average precision for the study case in the underground mine. All datasets, source code, and train test split are accessible at https://github.com/karana0103/EdgarObjDetection for future use cases.