An Emergency Landing Spot Detection Algorithm Based on Semantic Segmentation and Safety Evaluation

Due to emergencies such as GPS failures and hardware malfunctions, the UAVs may need to terminate their flight to prevent further damage. In this work, we propose an emergency landing spot detection algorithm that classifies terrain under the UAVs and identifies an optimal landing spot by safety evaluation. First, a knowledge distillation strategy is introduced for training a real-time semantic segmentation network DDRNet-23-slim with high accuracy. Next, the segmentation map is converted into a binary map and suitable landing areas are identified through morphology operations. Thin operation is then applied to represent the landing areas and the safety score at each point in the skeleton is estimated by the safety degree. Duplicate points are removed using a NMS algorithm, and a list of K points with top safety scores is generated and the optimum spot is then selected from this list. Both segmentation performance and safety gain of the optimum spot have been evaluated in our ZJLabid dataset. Experiment results prove that the proposed algorithm can effectively detect suitable landing spots and achieve a reliable and safe landing.