Vehicle Detection in Aerial Images Using Rotation-Invariant Cascaded Forest

Vehicle detection in aerial images has been taking great interest to researchers in recent years. It plays a crucial part in multidirectional applications, such as traffic surveillance, urban planning, and so on. However, the vehicle detection field faces many difficulties owing to the small size of the vehicles, different orientations, and the complex background. To solve this problem, this paper introduces a novel rotation-invariant vehicle detection method which is accurate, stable and has a simple structure compared with region-based convolutional network method. First, the data-driven method has been employed to generate the proposal region which will be applied for data augmentation. Second, this paper designs a method to obtain the rotation invariant descriptors by using radial gradient transform descriptors. Then, the rotation invariant descriptors are fed into the cascaded forest based on auto-context for feature learning and classification. The comprehensive experiments are conducted on the Munich vehicle dataset and UAVDT dataset. The results of experiment illustrate the satisfactory performance of the proposed method.