Dense Depth Maps Using Deep Learning with SLAM

Monocular SLAM systems are lightweight, fast, and well-suited for tracking the motion of fast-moving aerial vehicles. However, they are limited by their inability to recover absolute scale and by producing only sparse environmental maps. In contrast, deep learning-based depth prediction models can generate dense and detailed depth maps but are typically slower, temporally inconsistent, and sensitive to changes in viewpoint and domain. This thesis explores how these two modalities can be effectively combined to leverage their respective strengths. A method is proposed in which sparse 3D points from ORB-SLAM3 are used to estimate and correct the scale of predicted depth maps, enabling both scale-aware localization and dense 3D mapping. Evaluation shows that accurate scale correction is possible without model fine-tuning, although performance is highly dependent on the similarity between the deployment environment and the training data of the depth model. Additionally, integrating SLAM-derived points into the depth maps significantly improves their per-frame accuracy. However, when multiple views overlap, small inaccuracies in the SLAM points combined with residual errors in the predicted depth can result in noisy 3D environment maps. The system achieves real-time performance through several design choices, including parallel processing, resolution management, and limiting the number of frames for which depth is predicted.