Deep Convolutional Neural Network Based Vehicle Detection Methods On High Resolution Optical Remote Sensing Image

In high-resolution remote sensing images,vehicle detection is an indispensable technology in both civilian and military surveillance,e.g.,traffic management,urban planning,etc.Therefore,vehicle detection from remote sensing images has attracted significant attention worldwide.However,the current vehicle detection methods are mostly based on sliding-window search and handcrafted or shallow-learning-based features,having limited description capability and heavy computational costs.In the filed of computer vision,the deep learning based object detection methods have shown great performance in both speed and accuracy.But,directly utilizing them for vehicle detection in remote sensing imagery has many challenges: 1)Vehicles are relatively small in large scale remote sensing images,thus increasing the difficulty of localization.Moreover,due to the large size and wide view of the remote sensing images,their backgrounds are intricate,making accurate vehicle detection hard.2)In large-scale remote sensing images,near real-time vehicle detection speed is needed for many application.3)In remote sensing images captured from the top view,objects are rotated about the unit circle.The rotation angle of the object is useful information,but also makes accurate localization more challenging.4)The training data for vehicle detection in remote sensing images is much less,causing an over-fitting problem for deep convolutional neural networks(CNN)based methods.To address the above problems effectively,this thesis carried out research on vehicle detection in high-resolution optical remote sensing images based on deep learning.In this paper,we improve the deep CNN based object detection methods,aiming at improving both accuracy and speed of vehicle detection and estimating the orientation of vehicle during detection.The main work of this paper is as follows:1.We summarize two types of convolutional neural networks based object detection methods.By conducting vehcile detection experiments on Unmanned Aerial Vehicle(UAV)images,we conclude that the region proposal based CNN methods have higer detection accuracy and the regession based CNN methods have quicker speed.2.We propose a vehicle detection method for aerial imagery using region proposal based CNN and hard negative example mining.To solve the location and identification problem caused by large-sclae remote sensing images with complex backgrounds,we modify the region CNN based object detection method.Our goal is to improve accuracy.We propose a hyper region proposal network(HRPN),aiming at predicting all of the possible bounding-boxes of vehicle-like objects with high recall rate.Then,a cascade of boosted classifiers are used to further verify the detection results from HRPN,aiming at increasing detection precision.A large number of experiments show that this method can effectively improve the detection accuracy compared with the original CNN model.3.We propose a regression based CNN for arbitrary-oriented vehcile detection in aerial imagery.In order to estimate the orientation of vehicles in remote sensing images,we present an end-to-end regressin based single CNN to generate arbitrary-oriented detection results directly.Our method can detect vehicles using oriented bounding boxes with a simple pipeline,which is extremely fast.Numerous experiments on two open aerial image datasets have verified the effectiveness of the proposed method.