روشي نوين به منظور استخراج دقيق نقطه تقاطع شبكه‌ي راه در تقاطع‌هاي كانديد بر روي تصاوير بزرگ مقياس

A Novel Method for Accurate Extraction of Candidate Road Intersection Points from High Resolution Satellite Images

با توجه به اهمیت نقاط تقاطع شبكه راه در زمینه­های گوناگون مانند زمین مرجع سازی تصاویر ماهواره­ای یا برقراری ارتباط بین چند تصویر از یك منطقه­ی خاص و با توجه به دشواری و زمانبر بودن تعیین محل دقیق برخورد محور مركزی شبكه راه در محل تقاطع های كاندید در تصاویر بزرگ مقیاس توسط اپراتورها، در این مقاله سعی در ارائه روشی جهت تعیین موقعیت دقیق نقطه تقاطع محورهای مركزی شبكه راه در محل تقاطع شبكه راه در تصاویر بزرگ مقیاس ماهواره­ای گردیده است. در مرحله اول الگوریتم پیشنهادی با استفاده از  روش خوشه­بندی k-means و بهره­گیری از اطلاعات بافت زاویه­ای، پیكسل­های كلاس راه تعیین و در مرحله بعد با استفاده فیلتر Canny و  روش آنالیز لبه دایره مركزی ناحیه تقاطع، پیكسل­های لبه نزدیك به تقاطع شبكه راه تعیین و  در آخر با استفاده از روشی پیشنهادی، اقدام به برداری­سازی و تعیین محل دقیق نقاط تقاطع شبكه­ی راه شده است. تمركز این پژوهش، بر دقت بالای تعیین موقعیت تصویری نقاط تقاطع شبكه راه بطور مستقیم بوده و بر تعیین موقعیت تمامی نقاط تقاطع موجود در تصویر، تاكیدی نمی­باشد. پیاده­سازی روش پیشنهادی بر روی یك تصویر بزرگ مقیاس ماهواره­ای، موید توانایی آن در شناسایی محل تقاطع شبكه راه با دقتی در حدود نیم پیكسل می­باشد.

The increasing number of satellite sensors, with various characteristics in terms of spatial and spectral resolution, motivates the developments of automatic methods for better exploitation of their huge available information content. In this regards high resolution satellite images, which provide very detailed and accurate information of the urban areas, have attracted more research attentions. As an early step, geo-referencing of satellite images are essential by which image position observations can be transformed to the ground position information. This process needs some accurate control points which are traditionally provided by human operators. Considering the time-consuming and tedious task of control point provision automatic feature extraction from satellite images, especially distinct point features, has been the subject of many researches in the last decades. In this paper a novel method is proposed for accurate automatic extraction of road intersection points in candidate regions. These point features, if accurately extracted, are applicable for image geo-referencing or image-to-image registration. Roads are presented as wide ribbons in high resolution satellite images which cause a considerable uncertainty about the exact position of their intersection points. This uncertainty, which makes road intersection extraction a difficult task for human operators, motivates the aim of this research to develop an automatic and accurate extraction method. In the first step of the proposed method the candidate intersection region is divided to three clusters via a simple K-means clustering. The largest cluster is than labeled to background and omitted from the rest of processing. In continue angular texture information, extracted from a rectangular element, is used to detect road pixels. Road edge pixels are then found via the famous canny edge detector. Among the detected road edge pixels those which are close to the junction are found using Circle Centered Edge analysis. In the last step, a novel vectorization scheme is proposed to find out the exact position of road junctions. This scheme, developed based on the proximity and parallelism concepts, includes the following sub-steps: 1) line fitting to edge pixels here a line is fitted to the fist ten pixel and then the inclusion hypothesis of other pixels is verified by the evaluation of their distance to the fitted line, 2) eliminating the lines which are not related to road and determination of parallel lines, 3) grouping and merging of the same-type lines, 4) determining the co-side lines (i.e. road lines which are at the same side of a road segment) and 5) road centerline determination followed by positioning of the road intersection point. Given the road junction area, where at least 3 road segments intersect, the focus of this research is only on the exact and automatic positioning of the intersection point. The proposed method was implemented on four IKONOS sub-images taken from an urban area, Shiraz Iran, where in all cases the 0.5 accuracy proved the efficiency of the method. This method is also rotation invariant which means its applicability for all road segment orientations. Being resistant against the presence of vehicles and also the capability to extract more than one intersection points in a candidate region are other advantages of the proposed method.