Vehicle detection in remote sensing imagery based on salient information and local shape feature

Vehicle detection in high resolution optical imagery, with a variety of civil and military applications, has been widely studied. It is not an easy task since high resolution makes optical imagery complicated, which usually necessitates some rapid predetection methods followed by more accurate processes to accelerate the whole approach and to decrease false alarms. Given this “coarse to fine” strategy, we employ a new method to detect vehicles in remote sensing imagery. First, we convert the original panchromatic image into a “fake” hyperspectral form via a simple transformation, and predetect vehicles using a hyperspectral algorithm. Simple as it is, this transformation captures the salient information of vehicles, enhancing the separation between vehicle and clutter. Then to validate real vehicles from the predetected vehicle candidates, hypotheses for vehicles are generated using AdaBoost algorithm, with Haar-like feature serving as the local feature descriptor. This approach is tested on real optical panchromatic images as well as the simulated images extracted from hyperspectral images. The experiments indicate that the predetecting method is better than some existing methods such as principal component analysis based algorithm, Bayesian algorithm, etc. The whole process of our approach also performs well on the two types of data.