We propose an estimation method for the safety-level of local regions in aerial images for the emergency landing of Unmanned Aerial Vehicles (UAVs) based on Convolutional Neural Networks (CNNs), and introduce a new definition of safe areas and a new dataset. The estimation methods calculate scores of the safety-level for each region, and based on the results, the landing system detects safe areas where UAVs land without injuring humans, animals, buildings, artifacts, and themselves. Previous methods generally define natural flat regions, such as grass, lawn, soil and sand areas, as safe. However, if the flat regions are small and adjoin undesirable objects, the definition is dangerous and has the possibility of the injuring. Therefore, we introduce new definition to avoid the above complicated regions, and produce the dataset. Based on the dataset, we propose a CNN model to estimate scores of the safety-level. The proposed model can use various local and global features, and consider the environment of a target region. Hence, the proposed method estimates safe regions without the complicated ones, and then has better scores in the precision than the state-of-the-art method in experiments.
展开▼
