Heli-Tele Road Extraction from Helicopter Video

We present a learning based road likelihood computation method that uses aerial imagery and fuses information from several weak features. Our method is automatic, robust, and computationally feasible at the same time. We use road likelihood to align a colo

MITSUBISHIELECTRICRESEARCHLABORATORIES

http://

Heli-Tele:RoadExtractionfromHelicopter

FatihJieShaoandMaehara

DecemberWealearningbasedroadlikelihoodmethodthatusesaerialandfusesfromseveralweakmethodisautomatic,robust,andfeasibleatthesametime.Weusetoalignacolorhelicopterontoastreetmapto ndaddressofintheimage.IAPRConf.onMachineVisionMayThisworkmaynotbecopiedorreproducedinwholeorinpartforanycommercialpurpose.Permissiontocopyinwholeorinpartwithoutpaymentoffeeisgrantedfornonpro teducationalandresearchpurposesprovidedthatallsuchwholeorpartialcopiesincludethefollowing:anoticethatsuchcopyingisbypermissionofMitsubishiElectricResearchLaboratories,Inc.;anacknowledgmentoftheauthorsandindividualcontributionstothework;andallapplicableportionsofthecopyrightnotice.Copying,reproduction,orrepublishingforanyotherpurposeshallrequirealicensewithpaymentoffeetoMitsubishiElectricResearchLaboratories,Inc.Allrightsreserved.

cMitsubishiElectricResearchLaboratories,Inc.,2005Copyright

201Broadway,Cambridge,Massachusetts02139

We present a learning based road likelihood computation method that uses aerial imagery and fuses information from several weak features. Our method is automatic, robust, and computationally feasible at the same time. We use road likelihood to align a colo

MERCoLervPageSie2d

We present a learning based road likelihood computation method that uses aerial imagery and fuses information from several weak features. Our method is automatic, robust, and computationally feasible at the same time. We use road likelihood to align a colo

Heli-Tele: Road Extraction from Helicopter Video

Fatih Porikli, Jie Shao

Mitsubishi Electric Research Labs 201 Broadway, Cambridge, 02139, USA

Abstract

We present a learning based road likelihood computa-tion method that uses aerial imagery and fuses information from several weak features. Our method is automatic, robust, and computationally feasible at the same time. We use road likelihood to align a color heli-copter image onto a given street map to find address of buildings visible in the image.

1 Introduction

Alignment of aerial imagery and street maps is major challenge that geographic information systems are facing nowadays. In our setup, we want to determine the address of a chosen location in an image that is captured from a low-flying helicopter. One important application is auto-matic emergency services e.g. finding the address of a building in fire using aerial imagery. Although GPS in-formation is also available during the flight, it is often noisy with an off-set of 20 meters due to the motion of the helicopter and limited resolution of the GPS data. There-fore, additional refinement is necessary using the image and available street map using the only mutual features, roads, in both image and map as illustrated in Fig.1. Since extraction of roads is a time-consuming and it can not be performed manually for a real-time system, there is a need for automation.

Several approaches extract road candidates and then track roads [1]. One method models the context, such as shadows, cars, tree, etc. to improve the extraction of roads [2]. Learning methods were introduced as alternative automatic method by using grouping of parallel segments [3], detecting ridge-like descriptors using multi-scale methods [4]. Hough transform for the extraction of cross-ings [5]. Several methods make use of texture features. However, most of the existing approaches are either based on the hard heuristics and very specific to the type of the input data or not robust towards the various road and im-aging conditions exist in our application. Note that, it is sufficient to obtain road likelihood for each pixel, but not to precisely extract roads since such likelihood informa-tion is all is needed to align the input image to the street map.

To obtain a road likelihood map, we propose an auto-matic, robust, computationally feasible approach that uses low and high level image features. We selected features that provide most discriminating information by a learning based method, and tested several classifiers to achieve the accuracy and computational simplicity at the same time. In the following sections, we present these features and evaluation of classifiers, and sample results. Our system has already implemented as a part of our commercial ae-rial image analysis product.

Hide Maehara

Information Technology R&D Center 5-1-1-Ofuna, Kamakura, Kanagawa, Japan

Aligned Image and Map

Provided Street Map

Figure 1: Alignment requires extraction of roads.

2 Road Characteristics

There are several weak cues that indicate roads, how-ever, they are mostly not sufficient by themselves: Roads have salient edge features caused by lane divid-ers and intensity discontinuousness between buildings. Edges on the both sides constitute a pipe-line structure. Roads usually have homogeneous local orientation dis-tributions. Roads are continuous, so are contours. Width is almost constant and has an upper bound. Local curvature changes in a continuous manner except at cross-sections. Yet most roads are straight locally. Density of roads is proportional with the surrounding context. Road surface texture is different from buildings. Roads have a color range, i.e. they are not green or red.

These cues have advantages and drawbacks. Continu-ous contours with appropriate curvatures and parallel borders indicate roads, but they may fail when edges are occluded, shadowed, or simply not visible. While texture property may discriminate road r …… 此处隐藏：15850字，全部文档内容请下载后查看。喜欢就下载吧 ……