Georeferencing Algorithm for UAV Images Based on Semantic Segmentation

UAV image acquisition is a dynamic means of obtaining information from public areas. Based on UAV images, accurate positioning and rapid processing of public emergencies is becoming an important issue of concern for governments at all levels.This paper proposes a Georeferencing Algorithm for UAV Images Based on Semantic Segmentation(GeoSS). The GeoSS algorithm first imposes road semantic constraints on the reference map and extracts road feature points, uses the UAV image POS information to solve the image coverage area and obtains the road feature points in the corresponding area of the reference map. Then, the improved UNet semantic segmentation network is used to extract the road area in the UAV image.Experimental evidence suggests that the enhanced UNet architecture is capable of more efficient extraction of image features, and achieves 90.25% segmentation accuracy in the Semantic Drone Dataset, which is better than UNet. The segmentation accuracy of the baseline model has increased by 6.29%, and the mIoU indicator is 83.64%. On the self-built data set, the road semantic segmentation accuracy rate reaches 98.91%, and the mIoU indicator is 95.04%, which lays a solid foundation for subsequent feature matching based on road semantic constraints.In the feature matching experiment based on road semantic constraints, the SIFT algorithm was used to perform feature matching on the road constraint area, and the RANSAC algorithm was used to eliminate the mismatched feature point pairs, and the effective feature matching ratio between the UAV image and the reference image was increased to 57.9%. Ultimately, the geospatial coordinates of the UAV image were resolved by means of the cartographic correlation established between the pixel coordinates of the reference map and their corresponding geographic coordinates, culminating in the attainment of sub-meter level of georeferencing accuracy.