News

New method combines road segmentation and image-to-map registration

A crucial step in the process of vehicle detection from satellite images, is the segmentation of the road. The roads to be analysed are available as vector data, but to utilize this information the vector data and the images should be registered to each other. However, for most datasets the initial registration is not sufficiently good and has to be improved. On the other hand, to do the registration, keypoints like road positions need to be found. For the special purpose of doing vehicle detection along highways and country roads, we have therefore developed an automatic method that combines segmentation and registration. This is achieved by first using the vector data to sample a sufficiently large area in the satellite image around the road segments. The sampling is performed along profiles perpendicular to the direction of the road vectors, resulting in a long and narrow image (left figure). This image is then analysed line by line to find the most probable position of the road using a snake-like procedure. From the resulting segmented road (figure in the midle) and the original road vector, two distance images are produced (right figure) and the vector based image and the satellite image are aligned by using an image-to-image registration approach.

Road segmentation and image-to-map registration.

For more information contact: Arnt-Børre Salberg

Back to Section for Earth Observation