A Framework for Vision-Based Building Detection and Entering for Autonomous Delivery Drones

Autonomous delivery by aerial robots inside urban environments is getting closer to be operational everyday, but several concerns about proper navigation sources are to be addressed. As lot of GPS glitches are observed in urban environments, hereon a fully vision-based navigation framework is developed. The platform is assumed to have the ability to get into a vicinity of its destination using available "imperfect" navigation methods like GPS. Hereupon, the autonomous execution of mission is divided into four consecutive phases, all relying on a forward-looking camera: I) A novel vision-based method for detecting and confirming the target facade in the batch of neighboring buildings is introduced that performs matching using neural networks exploiting the texture features of facade segments; II) The intended window among the facade windows array is detected in real-time; III) The robot is guided to autonomously approach the intended window by a proper visual tracking algorithm; IV) Finally, a collision-free passage through the portal section of window based on visual Time-To-Contact estimation is commanded for the safe entrance. The network is trained using approximately 34,000 feature samples from a set of real-world building facades. As a result, a 95.4% accuracy along with 81.4% classification precision and 87.6% recall are achieved for the trained network in correct facade detection and confirmation. Also, the success rate of overall entrance mission is found to be 13 out of 15 in real world experiments, provided the initial distance being less than 20 meters.