A design and simulation of a target detection, tracking and localisation system for UAVs

Abstract

In computer vision multiple-object detection has gain significant interest by the researchers the last decade through the evolution in the field of deep learning. Nowadays, there are many architectures achieving great accuracy in detecting multiple objects in an image. On the other hand, tracking the detecting objects remains a very difficult task and still a lot of effort is provided in that field. In general, multiple-object detection, recognition and tracking are quite desired in many domains and applications. This paper presents a target detection, tracking and localisation solution for UAVs using optical cameras. A custom object detection model, based on YOLOv4-tiny, was developed based on YOLOv4-tiny and its performance was compared against YOLOv4-tiny and YOLOv4-608. While the target track algorithm in use is base on Deep SORT, providing state of the art tracking. The presented localisation method is capable of determining the position of ground targets, detected from the custom object detection model, with great accuracy. Finally, a guidance methodology is presented, responsible for creating real-time movement commands for the UAV to follow a selected target and provide coverage over him. The overall system was tested using Software-In-The-Loop (SITL) simulation in Gazebo with up to four UAVs.