Monocular Depth-Enhanced Visual SLAM for Autonomous Lawn Mowers

This thesis presents an approach to enable robust and efficient Simultaneous Localization and Mapping (SLAM) for autonomous lawn mowers by enhancing monocular visual SLAM with real-time depth estimation. Leveraging the ORB-SLAM3 framework as a foundation, the system integrates a deep learning based monocular depth estimator, FastDepth, to simulate RGB-D capabilities, overcoming the scale ambiguity and drift limitations commonly associated with monocular SLAM systems. The proposed solution is designed specifically for operation in garden environments, which are typically characterized by low-texture, repetitive patterns, and dynamic outdoor lighting, conditions that pose significant challenges to traditional feature-based SLAM algorithms. By fusing estimated depth maps with ORB feature tracking, the system generates scaled maps and improves trajectory accuracy, especially along feature-rich boundaries such as fences and pavements. The results of this study show that integrating depth estimation into a monocular SLAM pipeline significantly extends the applicability of vSLAM in low-cost robotic systems operating in complex, unstructured environments.