Automatic vehicle counts from QuickBird images


This paper propose a method for automatic vehicle detection in QuickBird images of small highways, with relatively low traffic density, frequent occurrence of tree shadows, and changes in illumination conditions. The vehicle detection is based on an elliptical Laplacian of Gaussian scale space methodology, where the vehicle locations are detected at local extrema in the image response to convolution with elliptical filters at various scales. The spatial extension of the vehicle candidate blobs are defined in a region growing step, and several object features are derived from spatial, spectral and contextual measurements of the image objects. Vehicle objects are separated from non-vehicle objects using a K-nearest-neighbor classifier, but different classifiers and features are applied to classify bright and dark vehicles. We also propose an approach for separation of dark vehicles and tree shadows. The overall performance of the vehicle detection processing chain is validated against manual vehicle counts, yielding a detection rate of 94.5% with only 6% false alarms. This detection rate may be considered acceptable for operational use in traffic monitoring.