Entropy-Aware Keyframe Selection for SemanticTopological Mapping

Abstract

One of the key challenges in autonomous mobile robot navigation is the construction of compact, semantically coherent topological maps. Current geometric simultaneous localization and mapping (SLAM) methods, including ORB-SLAM2, produce semantically empty and metrically accurate SLAM maps that are vulnerable to light changes, dynamic occlusion and perceptual aliasing. Conversely, vision-language foundation models trained on web-scale data encode rich semantic priors, yet their utility for online topological map construction remains insufficiently characterized. This paper presents a framework for entropy-aware keyframe selection in semantic topological mapping. We utilize Shannon entropy, computed over the latent embeddings of three state-of-the-art models (DINOv2, CLIP, and SigLIP), as an information-theoretic utility function to govern map node insertion. Extensive evaluations across four sequences of the KITTI Visual Odometry Benchmark (encompassing 9,504 frames) demonstrate that SigLIP achieves a high compression ratio (CR), yielding a 96.91% frame reduction on high-speed trajectories at the cost of larger spatial gaps between nodes. Conversely, DINOv2 yields the most consistent entropy profile across all environments (μ = 3.31– 5.72), while SigLIP drops to 2.19 on featureless highway sequences, indicating significantly reduced spatial discriminability under semantic redundancy. The results show that DINOv2 achieves the highest gap reliability on three of four sequences, SigLIP achieves the highest compression, and CLIP achieves the highest informativeness, giving insights for the design of perception systems in autonomous mobile robotics.