No abstract

No abstract

Most of the current research on autonomous landing of Unmanned Aerial Vehicles (UAVs) focus on rotorcrafts, which can fly horizontally over the landing site and then land vertically, with less landing risk. However, some of the unmanned fixed-wing aircrafts take off and land in the sliding mode of wheel landing gears, and their landing stage is difficult to control and easier to failure. In this paper, the autonomous landing of the fixed wing UAVs based on the visual navigation is studied. We propose a deep learning based Vision Transformer Particle Region-based Convolutional Neural Network (VitP-RCNN). Our VitP-RCNN uses Mobile Vision Transformers (MobileViT) as the backbone to accelerate feature extraction. Regarding the innovation of the present study, it is noticed that candidate boxes are chosen as the particles in our methodology, where the confidence is defined as the particle weight, the spatio-temopral correlation between adjacent image is used in video tracking and the Particle Filter theory is employed into the two-stage detection network to construct the present VitP-RCNN. Based on the spatial-temporal correlation, Our VitP-RCNN is efficient in predicting the particle state in the next frame through the state of the previous frame, so as to realize the continuous tracking of the landing marker. The experimental results show that on the Jetson AGX Xavier drone platform, our speed is up to 51FPS, enabling us to perform the landing marker tracking for the autonomous landing of the fixed-wing drones.INDEX TERMS Unmanned Aerial Vehicle (UAV); drone; autonomous landing; target detection; target tracking; deep learning