Deep Learning in Autonomous UAV Pursuit

Abstract

Unmanned aerial vehicles or UAVs have largely become and continue to be an inseparable part of modern warfare, security and surveillance systems, first aid response, aerial cinematography and many other sectors. Therefore, achieving full autonomy for UAVs and drones would ensure mass mobilization and utilization of these devices in large scale applications with more efficiency and precision without the need for deploying extensive human resources. A key aspect of autonomous flights is the capability of performing autonomous pursuits of target UAVs for military and civil purposes. Conventional autonomous pursuit algorithms rely on utilizing \lidar, radar and ultrasonic sensors or a combination and fusion of these devices. Each of theses sensors come with their disadvantages and shortcoming including a limited range of sight, massive data processing, expense and vulnerability to environment conditions. Thus, this research proposes a new sensor-free and vision-based algorithm for accomplishing fully autonomous UAV pursuit. This algorithm consists of two major parts including control and pose estimation. The flight controllers incorporate a digital four-axis proportional integral derivative (PID) framework and the pose estimator utilizes region-based convolutional neural networks (RCNN) for estimating a 3D bounding box over the target. The 3D bounding box keypoints are then extracted and combined with perspective-n-point (PnP) algorithm for estimating the precise relative pose of the target and pursuing it accordingly.