Detection Based Tracking of Unmanned Aerial Vehicles

Abstract

Object tracking is one of the fundamental problems of computer vision, which has many difficulties such as fast camera motion, occlusion and similar objects. Today, small and lightweight single board computers with very high processing power have been developed. Real-time processing of the computer vision applications on unmanned aerial vehicles has become possible with the integration of such single board computers within UAVs. In this study, a hybrid method is developed to detect and track UAVs by another UAV. A deep learning based approach which is one of the fastest and most accurate method in the literature, YOLOv3 and YOLOv3-Tiny (You Only Look Once), are utilized to detect the UAV at the beginning of the video and when tracking of the UAV is failed. Kernelized Correlation Filter (KCF) is used for real time tracking purpose of the detected UAVs. A dataset is created that consists different UAVs to train and test YOLOv3. Performance of the proposed methods are evaluated on this dataset.