A deep Learning-based approach for UAV's distance and height estimation through a monocular camera

Abstract

This paper offers a visual-based approach that detects the presence of Unmanned Aerial Vehicles (UAVs), or drones, and estimates their distance and height using a monocular camera. The proposed method consists of two fused models, the first model is the YOLOv4 objects detector which is trained on a new custom dataset to detect the UAVs presence and creates a box around the detected UAV followed by a deep feedforward neural network (DFNN)  model which is trained to take advantage of the box dimensions to estimate the UAV's distance and height. In other words, the estimation is based on the UAV's size and position in each frame. The trained YOLOv4 achieves a 99.79% mAP and an IoU of 79.61% and the proposed DFNN obtains an R2 score of 97.1% and an RMSE equal to 2.03m.