Navegação autônoma em ambientes internos baseada em estimativa de profundidade monocular

Abstract

Autonomous navigation in indoor environments constitutes an important challenge in mobile robotics, especially in applications that require low computational cost, real-time operation, and the use of embedded platforms. In this context, monocular depth estimation emerges as a promising alternative to traditional depth sensors, as it allows spatial information to be inferred from a single RGB image. This work investigates the feasibility of using modern monocular depth models in autonomous navigation systems, focusing on the FastDepth model due to its computational efficiency and suitability for resource-constrained devices. To this end, a comparative evaluation was initially carried out between the FastDepth, MiDaS Small, and Depth Anything Small models, considering metrics such as number of parameters, model size, inference time, and computational cost. Then, FastDepth was integrated into a reactive controller implemented in the Webots simulator, with the aim of evaluating its performance in different indoor scenarios. The system was tested in three distinct environments: a large room, a narrow corridor, and a room with tables, representing increasing levels of spatial complexity. The results demonstrated that FastDepth presents excellent performance in processing time; however, its limitations related to relative depth and frame-to-frame instability hinder its application in denser and less structured environments. It is concluded that reactive monocular navigation is viable in controlled scenarios, but still requires complementary strategies to ensure greater robustness in real and varied environments.