REKONSTRUKSI DAN TRANSFORMASI PERATAAN PERMUKAAN OBJEK 3D NON RIGID BERDASARKAN CITRA STEREO; NON RIGID 3D OBJECT RECONSTRUCTIONAND SURFACE FLATTENING TRANSFORMATION BASED ON STEREO IMAGE

HERMAWAN SYAHPUTRA, Agus Harjoko

2015 | Tesis | FMIPA

Recognition of non-rigid objects based on the image of object taken freely by using a regular camera is a difficult problem. The difficulty occurs because the view of object in the image can not retain the position of features in detail on non-rigid object, particularly on object with small observation area. This is due to non-rigid object has properties vulnerable to environmental changes. To overcome this problem, it is necessary to do the reconstruction and transformation of non-rigid surface of 3D objects. Reconstruction and surface flattening transformation of non-rigid 3D objects in a small observation area will result in quality of depth map/disparity image and object as a representative of the real object so it can improve accuracy of recognition. This study aims to reconstruction and surface flattening transformation of 3D objects based on the stereo image. The proposed stages are image acquisition using a stereo camera, preprocessing to change the image size, segmentation by removing the background of the object in the image, stereo image reconstruction based on segmented images, and 3D object's surface flattening transformation. Reconstruction of the disparity map is done by using the segmented image input. By using the SAD (Sum of Absolute Difference) matching algorithm and the window value and maximum disparity settings, the disparity map obtained from non-rigid image in small area that better than non segmented image input. To show the quality of the disparity image obtained, test processed by the recognition of non-rigid 3D object based on stereo images in some camera distances. The objects used are the leaves of plants. The result obtained in testing on object recognition based on this stereo image of leaf shows that the proposed recognition process is able to obtain an average accuracy of 82.7% for the three classes of plant varieties on camera distance of 30cm, 40cm and 50cm by using Gray level Coocurence Matrix feature extraction and Euclidean distance classifier. The result also shows that the camera distance to the object, ie, 30cm, 40cm and 50cm relatively has no effect on the recognition accuracy. New concepts and algorithms for the transformation of non-rigid 3D object surface flattening in the stereo image are also proposed. Result of the 3D object surface flattening process shows that the proposed method of transformation are well simulation and true concepts. Although the transformation of the object surface flattening concept is correct, the implementation of this flattening transformation method on leaf images taken from the stereo camera is not more than 70%. It is because the disparity image obtained in the reconstruction phase has not been able to show the depth of curvature of the object in detail according to the original object. The depth of the object curvature which is less accurate causes loss of intensity or texture are attached to the image coordinates during the transformation process of flattening

Kata Kunci : stereo; segmentation; disparity/depth maps; non rigid 3D object recognition; flattening transformation.