IMPROVING OBSTACLE RECOGNITION IN INDOOR ENVIRONMENTS FOR ROBOTIC SYSTEMS

Abstract

Accurate obstacle recognition plays a crucial role in enabling robotic systems to operate safely and effectively in indoor environments, especially in automotive-scale applications where safety and reliability is essential. This study focuses on enhancing segmentation performance for common indoor obstacles, including furniture and structural elements. The study presents a focused review and analysis of recent advancements in segmentation foundation models, with a focus on their potential in robotic systems and the use of architectures such as SAM2. To assess their applicability, we adapt a general-purpose segmentation model by fine-tuning it on RGB images from the NYU Depth V2 dataset, which contains a wide range of indoor scenes captured from viewpoints similar to those encountered by robotic systems, depth information was excluded to reflect the limitations of monocular camera setups typical in embedded robotics. The adapted model achieves improved segmentation accuracy, reaching 82.4% mIoU, and demonstrates greater robustness in recognizing obstacles within cluttered and visually complex scenes. Experimental results show consistent improvements over the baseline, with a 7.6% gain in performance, supporting the value of domain-specific adaptation for improving robotic perception. Although segmentation quality varies with the number of point-based prompts (performance is worse with 5 points than with 15) the improvement over the baseline in the low-input case is considerably higher, at a level of 27.9%, indicating strong suitability for systems where computational resources are limited. These results confirm the practicality of using pre-trained segmentation models, once properly adapted, to support the demands of obstacle recognition in indoor navigation for automotive-scale robotic platforms.