Abstract—Automatic vehicle detection from high-resolution remote sensing images plays a fundamental role in a wide range of applications. Various approaches have been proposed to address this issue in decades, however a fast and robust approach has not yet been found. It is still a very challenging task, due to the complex background, diverse colors and occlusions caused by buildings and trees. Traditional methods suffer from either high time complexity or low accuracy rate. To overcome the above shortcomings and consider the aforementioned difficulties, in this paper, we propose a simple and efficient approach to detect vehicles automatically. The proposed model lays emphasis on both speed and accuracy of vehicle detection, thus our proposed model consists of two stages: (1) To speed up localization, we apply the newly proposed Binary Normed Gradients (BING) to extract region proposals. (2) To enhance the robustness and improve the accuracy rate, we use Convolutional Neural Network (CNN), which combines feature extraction and classification. Therefore we modify the BING and design our own architecture of CNN to solve our problem. By comparing with start-of-the-art methods in extensive experiments, we demonstrate the effectiveness of the proposed approach in both speed and accuracy. Specifically, our method is more than 10 times faster than traditional methods, and our average accuracy is higher than state-of-the-art methods.

Index Terms—vehicle detection, Binary Normed Gradients (BING), Convolutional Neural Network (CNN), objectness

Cite: Shenquan Qu, Ying Wang, Gaofeng Meng, and Chunhong Pan, "Vehicle Detection in Satellite Images by Incorporating Objectness and Convolutional Neural Network," Journal of Industrial and Intelligent Information, Vol. 4, No. 2, pp. 158-162, March 2016. doi: 10.18178/jiii.4.2.158-162