This paper applies deep learning to obstacle identification in UAV flight. In order to achieve a better recognition effect, the dataset collates the two types of images of the simulation software screenshots and real-life photos. The basic principle of YOLO is briefly introduced. On the basis of not modifying the network, the data set is continuously trained using the pre-trained parameters. In this paper, we mainly optimize the loss function. The first is to optimize the difference between the simulation software screenshots and real photos by adding another weight. The second is class optimization, which increases the proportion of samples with low class confidence, like UAVs with different structures and colors. Finally achieved a good recognition effect.
Low-altitude Fixed-wing UAV Obstacle Recognition Based on Deep Learning
2018-08-01
254595 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Optical flow-based obstacle avoidance of a fixed-wing MAV
Emerald Group Publishing | 2011
|Optical flow-based obstacle avoidance of a fixed-wing MAV
Online Contents | 2011
|Flight Corridor Construction Method for Fixed-Wing UAV Obstacle Avoidance
Springer Verlag | 2023
|