OpenCV Implementation of Grid-based Vertical Safe Landing for UAV using YOLOv5

Abstract

The challenge of proving autonomous landing in practical situations is difficult and highly risky. Adopting autonomous landing algorithms substantially minimizes the probability of human-involved mishaps, which may enable the use of drones in populated metropolitan areas to their full potential. This paper proposes an Unmanned Aerial Vehicles (UAV) vertical safe landing & navigation pipeline that relies on lightweight computer vision modules, able to execute on the limited computational resources on-board a typical UAV. In this work, a grid-based mask technique is proposed for selecting the safe landing zones where each grid is parameterizable based on the size of the UAVs, which is implemented using OpenCV. A custom trained YOLOv5 model is the underlying building block for safe landing algorithm which is trained for aerial views of pedestrians, cars & bikes to identify as obstacles. The nearest obstacle-free zone algorithm is applied over the YOLOv5 output where boundary box locations are identified using Hue Saturation Value (HSV) filtering and then split into grids for safe landing zones where maximum coverage is taken into account while analyzing each scene. It performs a 2-level operation to prevent collisions while descending at different altitudes. Since UAV is expected to be processing only at predetermined altitudes, which will shorten the processing time, generating a PID signal for UAV actuators to navigate to the required safe zone with utmost safety and accuracy.