Online scene visibility estimation as a complement to SLAM in UAVs

Simultaneous localisation and mapping (SLAM) relies on low-cost on-board sensors such as cameras and inertial measurement units. It is crucial that the surroundings are visible to the cameras to maximise the accuracy of the system. An estimation strategy is proposed to augment ORB-SLAM2 that considers feature extraction capability, distribution of the extracted features in the image frame, and the ability of the algorithm to track features over time. The method is tested on challenging datasets, and the output is evaluated against different visibility conditions. The proposed method is shown to react appropriately and consistently to ‘less visible’ conditions such as fog, sunlight, and rapid motion in real time, with minimal computational load.