Multi-resolution mesh based 3D object recognition

Motivated by the need of the numerous real-world applications, 3D object recognition has become an active research field. The representation describes the sensed data and the object models, and it is a key issue in the recognition process, which decides the match strategy and the effectiveness and robustness of the recognition system. In this paper, we propose an improved 3D object representation first, which computes the local signatures of a given basis polygon on the surface mesh, and converts the signatures to a 2D array called the distance-angle (DA) images by weighted bilinear interpolation. This representation is adaptive to free-form objects and resistant to occlusion and clutter. Compared with the original representation, it has a more distinct meaning, easier operation, and adaptation to different resolutions and irregular triangle meshes. Secondly, based on the improved representation, a novel 3D recognition algorithm is presented, which has multiresolution mesh based, coarse-to-fine recognition. By matching the DA image of a polygon in the scene surface mesh with the DA images of models at low resolution, a model candidate set is obtained. The set is filtered in the neighborhood of the matched polygons in a high-resolution mesh and verified by the model candidate sets of other polygons. Experiments show that this algorithm needs less computation and is very accurate and robust