Integrating textural and geometric information for an automatic bridge detection system

We present some results on systems for automatically detecting bridges in high-resolution satellite images. We had made some preliminary explorations on the use of geometric models to detect bridges and round-abouts in panchromatic, high-resolution ((50 cm)² per pixel) satellite images. We had not yet systematically evaluated the system, but false alarms seemed to be its most important problem. In parallel, we had also built a system, which used local radiometric and textural features to classify terrain pixels into a number of semantically-meaningful classes, and then applied spatial relationships rules to these classes to detect and locate bridges. Validation showed a very low false alarm rate (around 5%), but also a low detection rate (around 40%).