Geoscene-based Vehicle Detection from Very-high-resolution Images

Vehicle detection is crucial to intelligent transportation system and other applications. Existing studies focus on detecting vehicles without considering the variations in geographic scenes, which have great impacts on the selection of image segmentation scales and detection algorithms. Moreover, they solely take advantage of visual features while ignore spatial distribution features. This paper presents a geoscene-based method to detect vehicles from very-high-resolution (VHR) remote sensing images. The method is implemented in three steps. First, geoscene segmentation and classification are made to obtain three categories of geoscenes so as to adopt different methods to detect vehicles for different categories of geoscenes. Second, the VHR image is segmented into image objects by using different scales based on different categories of scenes. The last step distinguishes vehicles from other types of image objects and optimizes the results by considering spatial distribution rules, which also vary from geoscene to geoscene. Experiments have been performed on Google Earth images and the final results are greatly improved and satisfactory.