Depth Estimation Networks for Navigation Safety in Indoor Robot Mapping

Abstract

The deployment of autonomous mobile robots in complex indoor environments relies heavily on robust perception systems to construct accurate maps and navigate without collisions. Monocular depth estimation networks have emerged as a prominent solution due to their low cost and spatial efficiency, yet their susceptibility to visual artifacts often compromises navigational safety. This paper presents a comprehensive simulation study focused on predicting navigation safety directly from the outputs of depth estimation networks during indoor robot mapping. By leveraging a highly realistic simulation environment, the research systematically evaluates how depth estimation errors correlate with imminent collision risks and mapping inaccuracies. A novel predictive framework is introduced, which analyzes spatial inconsistencies and confidence metrics generated by the neural network to forecast safety scores before the robot executes its trajectory. The methodology involves an extensive dataset generated within the simulation, capturing diverse lighting conditions, structural complexities, and dynamic obstacles. Analysis of the simulation results demonstrates that integrating a safety prediction module significantly reduces collision rates and improves overall mapping fidelity. This research bridges the gap between purely vision-based perception and safe motion planning, offering a scalable approach for evaluating and enhancing robotic navigation systems in a controlled, risk-free environment. The findings underscore the critical importance of uncertainty quantification in deep learning models applied to safety-critical domains.