Magnified Obstacle Avoidance and Landing Site Segmentation for Safe Path Generation of UAVs

Abstract

In order for drones to perform missions in areas close to people's lives, research is essential on technologies that can safely move away from obstacles and find and land safely on their own for complete missions. In this study, each element technology was manufactured and implemented based on simulation and embedded boards, and its performance was demonstrated through outdoor experiments. In order for drones to avoid collisions, they must detect obstacles with sensors, design safe margins to avoid collisions with obstacles, and create safe margins. In previous studies, a safe margin was designed to cover obstacles in a circle and furthermore, an ellipse. However, since these efforts are ultimately intended to design a safe margin that maintains the shape of the original obstacle. In this study, we conducted a study to design a safety margin while maintaining the original shape, and developed a collision avoidance algorithm that can be useful when maneuvering close to an obstacle such as in a city center. In addition, simulation using a matlab and outdoor experiments were conducted to prove the performance of the algorithm. In order to complete delivery to the island or to a space close to the lives of the people, it is essential to develop a technology for drones to find their own landing sites and land in a safe way. Existing research on segmentation and landing with camera images has been conducted at high altitudes. But this method narrowly predicted landing space due to shadows. Research using depth to find a flat area has a problem of judging a flat but non-landable area as a landing place, such as bush. Thus, this study conducted a research to find a landing place by supplementing each other's weaknesses using the semantic segmentation image based on camera image, deep learning, and depth map obtained by measuring LiDAR. In order to move to the landing site, when the landing zone is detected, the farthest area from the obstacle is found using the distance map, set as the landing spot, and move using visual geometry. In order to verify the performance of the algorithm, landing experiments were conducted in areas with bush and water, and performance was verified through actual landing.