Vehicle Detection and Classification in Aerial Images

Background Nowadays the problem of vehicle detection and classification on aerial images received from UAV (Unmanned Aerial Vehicles) has become important because of development of UAV technology and image analysis methods. Methods The paper describes a multilevel detection of vehicles and their subsequent classification. This method can be used for search of moving and stationary vehicles. In this work we propose new approaches such as: image segmentation into superpixels by SEEDS method, trainable five-level cascade detector of combined superpixels-regions, which uses technology of artificial neutral networks. Characteristics of regions are built based on their geometric and texture features (HoG and LPB descriptors) and directly from the image patch using technology of nonlinear autoencoders. Additional cascade of the detector uses data of moving objects in the image. Similar responses of the detector are combined and classified by color and type of the vehicle. Findings For training these algorithms a largest image dataset was compiled from different sources.The results of tests of detection and classification showed high accuracy. Improvements Algorithm is fast enough to allow on-board usage. The proposed method can be applied for road traffic monitoring, analysis of parking lots occupancy and other similar tasks.