Abstract

Obstacle avoidance is vital for unmanned aerial vehicles (UAVs). In this lab report, we present an approach for real-time metric depth estimation using unlabeled drone data. Initially, a pretrained depth estimator generates relative pseudo ground-truth depth maps. Consecutively, a lightweight model with real-time capabilities utilizes this training data to estimate depth maps up to a scaling factor that is subsequently eliminated through a final optmization scheme. Our evaluation explores various loss formulations, training approaches, and incorporation of temporal context, aiming to enhance baseline results. We improved on the baseline result and made our model runnable on a DJI Mini 3 Pro drone at up to 60 FPS. Moreover, we assess the capabilities of our scale estimation approach and the impact of different components. While our method proves efficient with ground-truth depths, it exhibits limitations when applied to imperfect predicted depth maps and requires further adjustments to be applicable in real-time scenarios