ティルトロータ型UAVの自動着陸のための安全な着陸地点の検出

抄録

In this paper, we construct a method of detecting safe landing sites for a tilt-rotor UAV from monocular images. We divide an image into small patches whose size corresponds to the area necessary for the UAV ’s landing, and calculate a safety score for each patch. The safety score is obtained by averaging two scores, the score from ground surface classification and the one from estimated flatness of ground surface. The ground surface in each patch is classified into 9 classes by using Convolutional Neural Network, and the flatness of ground surface in each patch is estimated by optical flows obtained from consecutive two images.