Research On Target Detection Technology Based On Aerial Image

Extracting object automaticlly from aerial and satellite images has been an intensive research topic in the fields of remote sensing and computer vision over last decades.This is because object detection plays a critical role in a diverse range of applications,such as urban monitoring,military reconnaissance,Border inspection,and estimation of population density.Detect object from aerial imagery has been one of the major challenges for computer vision scientists.There have been amount of works on car and pedestrian detection from aerial image,Though scientists developed a series of approaches for building extraction from optical remote sensing image,it is still remains a challenging task to develop generic and robust algorithms.One reason is that the images used often differ in terms of lighting conditions,quality and resolution.Another reason is that buildings may have diverse shapes and colors can be easily confused with similar objects such as cars,roads,and courtyards.So this paper want to propose a robust framework for building extraction in aerial images.The main research contents of this dissertation are as follow.First,this paperanalyzes the commonly used object detection algorithms.Random ferns classifier using 2bitBP description is adopted in this thesis,and random ferns classifier is used to object detection in the aerial video through the sample collection,classifier training,classifier test and other steps.After that,this thesis introduces salient object detection,Experiments show the merits and defects of ten classical salient object detection models in the Achanta dataset.Second,salient object detection models exploit background prior andcontract prior to attain state of the art results,but those methods fail when the background is complex.Instead of using background cues,this paper estimates the foreground regions in an image using objectness proposalsto capture super-pixels containing the object.Next,using the foreground connectivity measure,assigning foreground weights to super pixels.This thesis uses saliency optimization technique to combine our foreground weights with background measure to obtain smooth and accurate saliency maps.This paper compares the results of proposed approach with recent four methods:Saliency Filter(SF),GeodesicSaliency(GS),Manifold Ranking(MR)and Saliency Optimization(SO).Experiments show that the proposed approach performs better than the state of the art methods.Third,the proposed approach is applicable to buildings with saliency characteristics,but a major problem of proposed approch is that it has no capability to detect multi-building in aerial image.So this paper established a new general semi-automatic building rooftop extration method based on multi-scale image segmentation and model matching.In order to extract small and simple rectilinear rooftops from its background,this thesisusesseed region grow segmentation or localize multi-scale object oriented segmentation;and then usethe pose clusteringto delineate the precise position of complex rooftop,model matching techniques based on node graph search can find the correct building rooftop shape.Integration of these two methods makes extraction of buildings from simple rectangle rooftop to complicated building more practical.Finally,this paper extensicely evaluated methodsonsome aeria images.This article divided two experiments to verify detection algorithms.One group use the propoed salient object detection method for the saliency buildings,the other usemulti-scale image segmentation and model matching method for arbitrary buildings.Experiment results confirm that two approaches can correctly detect,locate,and extracte buildings,which demonstrates that two methods signifianctly outperforms the state-of-the-art.