3D-SLAM Implementation on Commercial UAV: Challenges and Future Insights

In recent years, the field of programming Unmanned Arial Vehicles (UAVs) has gained significant attention from researchers due to their substantial potential in various applications, including surveillance, inspection, and critical situations like examining buildings that are burning or collapsing. For this purpose, drones that have several sensors installed would be of good use in constructing a 3D map and localizing most of the objects within a certain area or room, which is also referred to as Simultaneous Localization and Mapping (SLAM). However, installing these sensors would harden the mission of the drone since it would mean more power consumption, more computations, and less navigation flexibility. For this reason, monocular visual SLAM has become the trend, which refers to using a sole camera to build the map and locate the objects in each scene. This approach introduces new challenges, one of the most crucial challenges is estimating the depth (i.e., distances within an image) of each scene from a 2D image. For this task, Deep Learning (DL) models have been considered as a solution for this problem, and with the continuous development in DL and the computational resources that can carry out the expensive training of DL models, it was shown that a 3D map reconstruction is possible utilizing 2D images. This work investigates the performance of a combination of different SLAM and depth estimation models implemented on a commercial drone. The main goal is to carry out a comparison between different methodologies of depth estimation that support monocular 3D SLAM algorithms.