System Development for Estimating Geolocation, Direction, and Velocity of Moving Objects in UAV Applications Using Monocular Camera

Unmanned Aerial Vehicles (UAVs) have become widely used across various fields, leading to an escalation in workload for UAV operators. Both computer vision and machine learning can significantly contribute to automating UAV operations. In UAV applications, e.g., traffic monitoring, object geolocation, direction, and velocity prove to be crucial data for effective UAV ground operations. This work introduces a system designed to detect and localize moving objects in UAV videos, using monocular camera. The proposed system is composed of four main parts: object detection, object tracking, object-depth estimation, and conversion of geolocation, direction, and velocity. Aiming for on-board implementation with limited resources, the selected algorithms prioritize processing speed over accuracy. For the first two main parts, the system utilizes YOLO and Lukas-Kanade algorithms respectively. It also employs k-nearest neighbors and polynomial regression along with MiDaS algorithm for object-depth inference models. The last part involves transforming position from the camera plane to the real world plane. Following testing of the proposed system at two UAV heights, it was found that the MiDaS with polynomial regression model performs better in geolocation and velocity estimation, al-though it exhibits slightly lower accuracy in estimating direction. This presents our initial step in developing the proposed system, with plans for system enhancement in the future.