Vehicle Detection in Satellite Images for Development of Traffic Statistics Title Sattrafikk -vehicle Detection in Satellite Images for Development of Traffic Statistics – Project Results 2007 Executive Summary

The road network is a key resource of enormous value for the nation. A satellite image is covering a large area instantaneously and can thereby be a source of road traffic information in a snapshot of time. Manual vehicle counting is not realistic, but automatic image analysis methodology based on pattern recognition is a promising alternative. The “Road Traffic Snapshot” project (2006‐2007) showed that such a future system for automated road traffic counts is feasible. The follow‐on project SatTrafikk has improved the methodology and verified it on a far larger set of satellite images. Based on advice from the road authorities of Norway, we have selected a set of study sites from different parts of the country, such that our image data represents the diversity of road types and solar illumination conditions. Road and vegetation masks are applied to the image so that the search for vehicles is restricted to the (paved parts of the) roads only. For segmentation, we have applied techniques that seek to locate the modes of the image histogram. The resulting segments are then examined by feature extraction and classified adopting the maximum likelihood method. Additionally, we propose a new approach for car shadow removal. The described methods was implemented and tested against manual vehicle counts. We also compared the results to traffic statistics estimated from single‐point measurements. The majority (90%) of vehicles that are found in the segmentation step are correctly classified as vehicles. Comparing the number of cars found in the image to the number of cars that according to single‐point measurements are expected to be found in the image, the manual counts had accuracy between 79 % and 120%, while the automatic counts had accuracy between 68% and 157%. Some shadows and road marks are hard to distinguish from vehicles. Some vehicles have low contrast and are not captured in the segmentation step.