Survey and Implementation of Passive SLAM Techniques for Natural Environments

Abstract

Autonomous robots, self-driving cars and delivery drones which are viewed as future technological advancements, depend on a specialized class of framework known as Simultaneous Localization and Mapping (SLAM) to enable autonomous navigation. SLAM helps an autonomous vehicle to chart a course through unknown terrain by building a map in real time from features detected by its onboard sensors. However, these frameworks remain under continual refinement to achieve true autonomy, as sensor limitations and computational constraints can occasionally limit the system’s capability to capture the critical details from the environment. The challenge becomes even more pronounced when relying solely on passive sensors such as cameras and inertial measurement units (IMUs). Sometimes, these sensors have to be set in power restricted settings, or stealth mode applications, where active sensing could cause harm to living creatures. This thesis aims to present an extensive study of the diverse sensor modalities employed in mobile robotics for navigation in rural environments, including forested and natural terrains. It encompasses advances in SLAM methodologies that incorporate both passive sensors (e.g. stereo cameras, IMUs, monocular cameras and RGB sensors) and active sensors (e.g. 2D LiDAR, 3D LiDAR and ultrasonic sensors). We then implement the leading passive-based SLAM approaches like ORB-SLAM2 and ORB-SLAM3 to evaluate their performance under realistic conditions in the forest environment. The results indicate that while the average Absolute Trajectory Error (ATE) in indoor environments remains within approximately 10 cm, it increases significantly to 20–30 meters in forest scenarios, demonstrating a clear performance gap. Despite strong reputation, these SLAM algorithms show room for improvement. Choosing better sensors can further enhance their performance. This thesis proposes future research directions, highlighting the potential of multisensor fusion. In particular, integrating stereo cameras and thermal imaging with learning-based SLAM frameworks can achieve more reliable and resilient mapping in natural environments.