Multi-Modal Place Recognition and Pose Estimation for Autonomous Rovers in Unstructured Environments

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Electrical Engineering | Master's thesis

Department

Major/Subject

Mcode

Language

en

Pages

122

Series

Abstract

Autonomous navigation in planetary-like environments presents unique challenges due to the absence of GPS signals, limited semantic structure, and visual ambiguity caused by repetitive textures. Traditional place recognition and localization methods either rely on dense maps and structured environments or only provide place retrieval without estimating full 6-DoF (Degrees of Freedom) poses. This limits their applicability in the context of real-time Simultaneous Lo-calization and Mapping (SLAM) for planetary exploration. This thesis addresses the problem by developing a multi-modal system that per-forms both place recognition and relative pose estimation in unstructured environ-ments. The proposed approach fuses transformer-based visual features from DI-NOv2 with LiDAR-derived 3D descriptors from SONATA, aligning them in 3D space to produce interpretable correspondences. Place retrieval is performed by ag-gregating DINOv2 descriptors with SALAD and searching via FAISS indexing. The method is evaluated on the Etna volcano dataset, representative of planetary ter-rains. Results show superior retrieval performance over NetVLAD and TransVPR and more stable pose estimation than handcrafted or regression-based base-lines. Multi-modal fusion improved robustness in low-texture and low-light scenes, supporting the value of combining vision and LiDAR. The system produces interpretable outputs and operates within real-time SLAM constraints for retrieval, although further optimization is needed for pose estimation. These findings high-light the feasibility of delivering explainable 6D poses for SLAM in extreme envi-ronments, with potential applications in exploration, disaster response, and agri-culture.

Description

Supervisor

Zhou, Quan

Thesis advisor

Jensfelt, Patric
Folkesson, John

Other note

Citation