A Real-Time Monocular Vision-Aided Target Detection and Tracking for Micro UAV in GPS Denied Environment

Abstract

Surveillance and tracking in GPS denied environment is an important application for vision-based micro aerial vehicles (MAV). In this paper, we developed an autonomous MAV with embedded onboard vision processing as a navigation system, which can achieve fast and consistent flight to pass through targets in GPS denied environment. It is very suitable for tracking the targets with specific marks in GPS denied environment. The main advantages of it are the ability to calculate the vision data onboard in real time at a weight of only 348 g and the quickness and consistency of the movement compared with the existing similar MAVs. In this method, we designed two alternative visual schemes to process the vision data. In the first scheme, an Openmv machine vision platform is used and for another scheme, a lightweight deep learning platform, named Huawei Hisilicon 3516 and the YOLO model after pruned are used. The target information is sent to the flight controller under the Mavlink protocol. Predictive control scheme with longitude and lateral decoupled is designed to stabilize the MAV in a high-speed maneuver. Finally, the flight test demonstrates that the solution can make the drone pass through stationary or moving circular and rectangular targets accurately autonomously.