An Appearance-Based Approach for Indoor Environment Semantic Classification for Autonomous Navigation

• Alternei de Souza Brito UFAM
• Emanuelle de Souza Gil UFAM
• Felipe Gomes de Oliveira UFAM

Resumo

Autonomous ground vehicles operating in indoor environments must accurately perceive and classify their surroundings to ensure safe and efficient navigation. A fundamental step in this process is the semantic classification of environments, which provides contextual awareness to decision-making modules. In this work, we propose an appearance-based indoor environment semantic classification process that leverages DINOv2, a self-supervised vision transformer, for feature extraction. The extracted features are used to identify the type of indoor environment such as corridors, offices, or storage rooms based on visual features. To evaluate the effectiveness of our approach, we trained and tested eight different state-of-the-art deep learning models to validate the proposed semantic classification pipeline. Experiments were conducted on the KTH-IDOL2 dataset, comprising diverse indoor scenes captured from the perspective of mobile ground robots. Results demonstrate that our approach based on DINOv2 produced the best overall performance, with ConvFormer obtaining the highest classification accuracy, followed closely by DenseNet201. The proposed pipeline demonstrates to be a promising strategy for enhancing the perception capabilities of autonomous systems operating in structured indoor environments. Additionally, experiments evaluating images under varying temporal and lighting conditions were conducted, demonstrating the robustness of the proposed approach.

Palavras-chave: Training, Visualization, Accuracy, Semantics, Pipelines, Lighting, Feature extraction, Robustness, Indoor environment, Robots, Semantic Environment Classification, Indoor Scene Understanding, Domain Generalization, DINOv2, Au-tonomous Navigation