Vehicle Detection From Aerial Photographing Under Different Illumination and Pose

To solve the problem of low detection accuracy of vehicle detection from aerial photographing under different lighting conditions and different postures, a new method based on the Fourier-HOG algorithm was proposed. This method was based on a sliding-window detection approach. First, image preprocessing, which selectively removed the background region, greatly improved the efficiency of detection and reduced the false alarm rate. Second, illumination invariant features were extracted based on local sensitive histogram and then the rotation invariant Fourier-HOG features were extracted. Finally, from the above features, the vehicle and non-vehicle were discriminated in a linear support vector machine (SVM) classifier. For post-processing, nonmaximum suppression technique was used to reduce a target multiple-detection. Results of the proposed vehicle detection on the Google Map dataset show that it has a higher degree of detection accuracy and consumes less time than that of the original Fourier-HOG detection method. Therefore, this method is a valid vehicle detection from aerial photographing.