Fusão Multimodal via Atenção Cruzada Bidirecional para Mapeamento Semântico de Ambientes Internos

Abstract

Semantic mapping is fundamental for mobile robots, as it provides a spatial representation associated with semantic labels that support navigation and context-aware decision-making. This dissertation addresses indoor semantic mapping by treating it as supervised scene classification from RGB images and estimated monocular depth. We propose a multimodal architecture based on self-supervised representations and Bidirectional Cross-Attention fusion between appearance and geometry. The approach is evaluated on KTH-IDOL2 and COLD under protocols of temporal generalization, illumination variation and interpolation, and cross-domain transfer. Results show high performance and consistent gains from multimodal fusion in challenging scenarios, as well as greater spatial coherence in the 2D semantic maps generated from the predictions. This strategy increases the reliability of perception for mobile robots in real indoor environments, supporting navigation and decision-making under photometric changes and domain shifts, without requiring additional sensors beyond an RGB camera.