Benchmarking SLAM and supervised learning methods in challenging real-world environments

For an autonomous robot to navigate in unknown environments and interact with its surroundings, it must accurately estimate its spatial position and independently identify perceived objects. To achieve this, Simultaneous Localisation and Mapping (SLAM) is applied in conjunction with supervised learning techniques for object detection and classification. Given the wide range of open-source solutions, this thesis aims to find approaches best suited for complex real-world conditions and benchmark their performance. The sensor data for the evaluation was taken from the publicly available EnvoDat database, which was created specifically for this purpose. The benchmark investigates LiDAR-based (FAST-LIO 2, GLIM, DLIO, HDL-SLAM) and visual-based (RTAB-Map) 3D SLAM approaches and pre-trained supervised learning models (YOLOv8 and YOLOv12). As shown in the results, LIO (LiDAR-Inertial Odometry) methods generally deliver superior and consistent performance in all environments. However, the effectiveness of the evaluated training models declines significantly under challenging lightning conditions, as reflected by a drop in mean average precision.