Vehicle Detection in Aerial Images Using Selective Search with a Simple Deep Learning Based Combination Classifier

Contrary to the growing zest for bringing in place a complex detection methodology, aiming at improvement in performance of existing methodologies for detecting vehicles in aerial images, this novel piece of work sets forward a much simpler approach with superior results. It was found that methods that showed exemplary performance on common benchmark datasets otherwise, their performance dropped remarkably on aerial images. To achieve performance at par or comparable with the state-of-the-art methods on common benchmark datasets, several adaptations have been suggested in literature to existing methods, for detecting small vehicle instances in aerial images. This ranges from adaptations to the object proposal methods to introduction of more complex deep learning-based classifiers such as fast RCNN and faster RCNN. However, these methods have their own limitations along with the growing increase in system complexity. In this work, a novice, simple and accurate method has been proposed for the detection of small vehicles from aerial images. The experiments have been performed on the publicly accessible and diverse Vehicle Detection in Aerial Imagery (VEDAI) database. This novice technique utilizes Selective Search algorithm as the object proposal method in combination with a deep learning-based framework for classification, which comprises of a simple Convolutional Neural Network (CNN) architecture proposed in combination with a simple Deep Neural Network (DNN) architecture. The DNN utilizes Histogram of Oriented Gradients (HoG) feature input to generate output features that combine with the CNN feature map for final classification. This method is much simpler and achieves a significant accuracy of 96% in vehicle detection, which is much superior to any of the methods tried for aerial images in literature so far.