Robust detection and localisation of cars in airborne imagery

Abstract

With the ever-increasing demand in the analysis and understanding of aerial images, this work is focused on the use of implicit models for the detection and localisation of small targets in static airborne imagery of typical urban scenes with associated ground sampling distance of few centimetres. These scenes are characterised by the existence of complex structures and heavy clutter. More specifically, it focuses on the detection of non-occluded cars as it has many useful applications in security, surveillance, remote sensing and other domains. This is a tremendously-challenging problem because of the huge inter-class similarity among car and other non-car targets in urban scenes. Nevertheless, it will be demonstrated that high precision rates can be achieved using robust ensembles of image descriptors alongside linear classification techniques, e.g., linear Support Vector Machines. The original contributions of this work include: i) novel methods to reduce search areas by evaluating the likeliness of detection windows to contain targets, (ii) accurate methods to estimate the dominant orientation of detection windows, (iii) robust ensembles of image descriptors of low dimensionality relative to the state of the art that depict the distributions of gradients, texture, colours, spectral information and the second-order statistics of each and (iv) a new method to extract regions that are placed around roads using a Gaussian Mixture Model classifier. The performance of the proposed frameworks was evaluated against two benchmark datasets of static images, namely, the Vaihingen and the Overhead Imagery Research datasets. Results show that the proposed frameworks are superior to and outperform the state of the art.