Vehicle Detection Based On Optical Principle In Optical Aerial Image

Optical object detection,which has close relevance to the physical optics,is a fundamental research direction of computer vision.As the technology of UAVs and optical imaging develop rapidly,UAVs have been broadly applied to many areas such as electronics,rescue,patrol and transportation.Many scenes request UAVs to be capable of automatically searching and monitoring the interested targets.As one of the key technologies,Vehicle detection technologies based on UAVs is becoming a research hotspot.What this thesis studies is a subject combined with science and engineering.The difficulties and problems of vehicle detection in optical aerial images are as follows:1.Existing optical images datasets are not complete.Object detection datasets only have large-scale objects and have no overhead perspective images,while UAV datasets only have vertical perspective images.2.Traditional methods that deal with vehicle detection in aerial images have certain limitations.Road detection is only for vertical angle of view.Moving object detection cannot detect static objects.Meanwhile,it is difficult for the emerging deep learning methods to detect small-scale objects.3.Vehicles have various viewpoints and small scale in optical aerial images.These targets are hard to detect and easily general false alarm during the vehicle detection.To solve these problems,this thesis proposed a coarse-to-fine vehicle detection framework that consists of a region proposal method and an eliminating false alarm method.The main studies and contributions of this thesis are:1.Aiming at the problem that public datasets poorly reflect reality,a lot of aerial images have been gathered and a new UAV vehicle dataset is built in order to study the vehicle detection under the platform of UAVs.2.As a resolution to the general vehicle detection method in optical aerial images,aggregated channel feature is chose to be the region proposal method.The represent performance of optical features is explored and multiscale features have been studies.Fast feature pyramid is used to reduce the detection time by approximating the features.To reduce the training time and avoid overfitting to some extent,Ada Boost based on the decision tree is chosen as the classifier.Experiment shows that the region proposal method effectively extract vehicle targets from aerial images.3.For dealing with the high false alarm rate,this thesis adopts an eliminating false alarm method based on optical principle.Because in a uniform medium light travels in straight lines,there are shadows around vehicles.Absorption and diffuse reflection occur after light comes through windows of vehicles.As a result,the intensity of the reflective light is low from a distance.According to the characteristic of vehicles having dark region like shadow and windows,features of vehicles in dark channel images have been studied.Threshold segmentation and other approaches are utilized to distinguish the background environment and vehicle targets.Experiment shows that the proposed method can improve the vehicle detection performance.