كليدواژه :
ابرنقطه , كلاسه‌بندي , قطعه‌بندي , توصيفگر
چكيده فارسي :
امروزه پردازش خودكار ابرنقاط ازجمله موضوعات مهم و پرچالش در فتوگرامتري و سنجش‌ازدور مي‌باشد. لايدار به‌عنوان يك سنجنده فعال توانايي اخذ مستقيم ابرنقطه داراي مختصات سه‌بعدي با دقت بالا را دارد. با گسترش تكنولوژي و نرم‌افزارهاي پردازش تصوير امكان توليد ابرنقاط با دقت بالا براساس تناظريابي چگال از مناطق همپوشاني تصاوير هوايي نيز فراهم گشته است. پردازش‌هاي مربوط به ابرنقاط نظير قطعه‌بندي و كلاسه‌بندي عموماً داراي هزينه محاسباتي بالايي بوده و زمان‌بر مي‌باشند. ازاين‌رو ارائه روندي كاربردي كه بتواند با سرعت پردازش بالا به دقت مناسبي دست يابد، همواره مطلوب كارشناسان بوده است. در اين مقاله روندي با رويكردي متفاوت جهت قطعه‌بندي ابرنقاط مطرح شد و سپس با بهره‌گيري از مفهوم شي‌ءگرايي روندي براي كلاسه‌بندي قطعات شناسايي شده، ارائه گشت. در اين راستا، ابتدا تراكم ابرنقاط كاهش‌يافته و سپس قطعه‌بندي براساس گسترش ناحيه و با استفاده از ميزان انحنا و بردار نرمال صورت گرفت. با برچسب‌گذاري نقاط كنارگذاشته شده در مرحله كاهش تراكم براساس جستجوي دقيق اطراف نقاط قطعه‌بندي شده، نتيجه نهايي قطعه‌بندي حاصل گشت. در مرحله بعد براي قطعات شناسايي شده، توسيف‌گرهايي براساس ويژگي‌هاي هندسي و ساختاري عوارض مختلف معرفي و توليد شد. درنهايت نيز براي كلاسه‌بندي قطعات شناسايي‌شده از الگوريتم KNN استفاده شد. روند پيشنهادي در 6 ناحيه مطالعاتي پياده‌سازي شده و مورد ارزيابي قرار گرفت. ارزيابي نتايج دقت متوسط %42/91 براي شناسايي سه كلاس ساختمان، پوشش گياهي و سطح زمين را نشان داد كه حاكي از قدرت بالاي روند پيشنهادي است.
چكيده لاتين :
Nowadays, automatic point cloud processing is an important and challenging topic in photogrammetry and remote sensing. The LiDAR has the ability of collecting the accurate 3D point cloud from the earth surface, directly. Moreover, recent advances in image processing provide the capability of producing 3D point clouds with high accuracy using dense matching from the digital aerial images. The point cloud segmentation and classification algorithms are usually time consuming and have high computation cost. In this paper a difference object-based approach was proposed for point cloud classification. In this approach, at first the points were segmented into some regions; then these regions were classified into considered classes. In this regard, firstly some boxes with predefined side size were placed on point clouds and each box was analyzed separately. In order to reduce the point density, the points in each box were removed except the nearest point to the center. Then, the region growing algorithm was employed to segment the points with reduced density based on normal vector and curvature value of each point. Afterward, around of each segmented point was searched for labeling the remains points. In other words, the points which have normal vector close to considered point were labeled same as that point. After point segmentation, for each segment some potentially features were selected and produced in order to detect buildings, vegetation as well as grounds. The features should be selected accordance with the geometrical and structural characteristics of the objects. In this paper some features including mean curvature, area, perimeter, boundary irregularity, flatness, elevation, and being terrain or off- terrain were generated. The Alpha shape is a triangulation based algorithm which has the ability of reconstructing the object shape using a set of dense and irregular points. The Alpha value determines the level of details in the reconstructed shape. After computing the shape of the considered segment using Alpha shape algorithm, calculating the area and perimeter was feasible. In order to analyze the boundary irregularity of the segments, the ratio of area between two reconstructed Alpha shapes with two different Alpha values is computed. For each segment a plane was approximated using the MSAC algorithm and the ratio of points in that plane and out of that plane was computed as flatness value. The SMRF algorithm was employed for specifying the off- terrain points. The height of an off-terrain point was acquired by computing the difference between that point and the closest terrain point. Thus, for each segment a feature vector was obtained. Finally, some training data was collected and the segments were classified by KNN algorithm. The proposed approach was implemented and evaluated in 6 different test areas. Although the area 1, 2, 5 and 6 were acquired by LiDAR, the point density of area 1 and 2 is equal to 4 point per m2 and the point density of area 5 and 6 is equal to 65 points per m2. The area 3 and 4 were acquired by dense matching of digital aerial images and theirs average point density is equal to 20 points per m2. The accuracy of proposed approach in area 1 to 6 were 92.25%, 93.44%, 91.44%, 89.23% 92.46% and 89.73%, respectively. The evaluation results clarify the good performance of proposed approach in different areas with various land covers and point densities.
