[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
We propose a method to retrieve range images from database robust for data loss.
It becomes easier to get range images by laser range finders. A range image has three-dimensional position data about points on surfaces of objects, and the position data are represented in the coordinate system defined by the range finder. Range images have several merits. (1)they can represent three-dimensional information directly and precisely. (2)they can make geometrical registration easily by the manipulation such as rotation and translation. (3)the values of position data are not affected by the environment. These merits promote researches to use range images for object recognition and pose estimation. In these researches, there are many range images of an object captured from various directions. Hence, range image database which have a lot of range images of various objects captured from various directions will be widely used in near future. On the other hand, there have been no research that proposes the method to retrieve range images from the database using a range image.
In the range image retrieval, it becomes a serious problem that a range image includes data loss which occurs when we get the range image. The reasons for the data loss are difference of the capturing angles, out of place from the measurable range of the range finder, and the properties of the range finder's sensor. These data loss may make it hard to identify range images of the same object. In our research, we assume that these data loss exist not only in the input data but also in the data of the database. We can not retrieve range images neither by comparing range images in the feature space, as to be done in the content based image retrieval, nor by matching range images using feature points, as to be done in the object recognition, because of the difference of the rates and the positions of the data loss among range images . For these reasons, the retrieval method has to be robust for data loss. In our method, we emphasis on the common part between the range images. By making use of the common part, we can retrieve range images robust for the data loss. Since this method costs much time, we also discuss the reduction of the calculation time.
Our method has three stages. At the first stage, we generate new range images by rotating the input range image by the constant angles. At the second stage, by shifting and overlapping the generated range images to the range images in database, we detect the proper common part between them. The common part is determined by the depth difference in the same grid of the compared two range images. At the last stage, we get the evaluation value by calculating the percentage of the number of the grids in the common part to that of input data.
This method takes a lot of time. The reasons are the large amounts of calculation in the registration, and in the evaluation. In the registration process, the proper position that means the position where the common part is maximum is searched. To reduce calculation time, the step of searching the proper position is enlarged. In the evaluation process, we choose the comparison candidates before calculating the evaluation value for all range images in the database. The method of choosing the candidates is as follows. At first, we calculate the evaluation values among all range images in database. Second, we choose representative range images in the database. Third, we calculate the evaluation values between the input range image and the representative range images. The range images whose evaluation values for the representative range images are approximately equal to those of the input image are chosen as the candidates. By this way, the number of calculating the evaluation values between two range images may reduce to the sum of the number of the representative range images and the candidates.
This paper reports the effectiveness of the retrieval that use the correlation between the range images by conducting to experiments in range images retrieval. As for choosing candidates, it can keep the good precision and reduce calculation time, when the loss caused by the out of range does not exist.