In recent AR applications, not a barcode-type marker but also the texture marker is used for. In this paper, we proposed a new matching algorithm for texture marker recognition. Conventionally, the keypoints based matching algorithm such as SIFT is used for the texture marker recognition. However, it causes the position errors because the accuracy of position is affected from illumination change and noise. In order to solve the problem, we propose two-stages matching algorithm. First, the conventional keypoints based matching is used for initial guess of camera pose. Secondly, the edge based matching is applied. The search of edge correspondences is performed in a local search area iteratively. Furthermore, in order to evaluate the estimated pose, we proposed to compare the estimated pose and the reference obtained from the mechanical 3D probe sensor. Experimental results show that the proposed method can improve the accuracy of camera pose and the computational time is reasonable.
This paper describes an improvement of object pose estimation method for bin-picking problem. A bin-picking system by using industrial robots contributes to factory automization. Industrial robots need to recognize the poses of the object in a bin because they have different poses. We propose a pose estimation method based on point pair feature (PPF) matching. PPF is pose-invariant and has high robustness of occlusion situation. PPF matching performs well in easy bin scenes. However, the successful rate of the pose estimation by PPF matching often decreases in some difficult scenes such as a bin with shiny or black objects because these objects cause bad measurement data with sparse point clouds and low signal-to-noise ratio. Our proposal evaluates PPF reliability by using PPF correspondence between scene data and object database. A PPF which votes for the same pose several times causes false pose generation. Thus, we judge such PPFs unreliable. By rejection of poses generated by using unreliable PPFs, successful rate of pose estimation increased 45% in a difficult bin scene with shiny and black objects. Experimental results show effectiveness of our proposal.
These days, carsharing service which is a service to share cars with multiple persons has attracted attention. Carsharing service aims at promoting the use of public transportation facilities and suppressing the excessive dependence on the automobile. Users of the service can use the cars immediately just after the reservations through their mobile terminal once the registration of each user is conducted on the first use. The users can use the cars stress-free because there is no procedure in renting and returning and can use for a short period of time such as about 15 minutes. However, there is a problem that responsible for an accident or malfunction cannot be clear because much of the service bases are operated without field staff. Therefore, we have developed automatic visual inspection system to be used for sharing cars with multiple persons. Our inspection apparatus consists of a long LED bar light and a camera. This system automatically inspects a side surface of the car that is going across in front of the apparatus. In an experiment, inspection accuracy of scratches and dents on car body surface has achieved 95.6 %. Our system is able to identify who is responsible for an accident by comparing inspection results before and after use. Furthermore, our system has a possible to lead to the cost reduction in car rental business.
We propose a method to observe cardiac beat from 3-D shape reconstruction by using the grid-based active stereo. In this study, we report preliminary experiments to evaluate validities of the proposed method. As the result of comparing experiments by our method and electrocardiogram (ECG), we confirm sufficient correspondences between the two peak intervals. Our method realizes the separate extraction of cardiac beat and respiratory movement. We try the visualization of cardiac beat by using the color mapping of the depth displacement plotted on the 3-D surface rendering of chest. And, the depth change and the spatial phase difference caused by cardiac pulsation is found on the chest region.
This paper proposes a fusion technique that takes noisy low-resolution distance information and high-resolution color information, and produces accurate and high-resolution 3D measurements. In our method, first, tangent planes on each superpixel of the high-resolution color information are estimated from the low-resolution distance information. Then, using ray-tracing calculation, these tangent planes are divided into groups consisting of smooth-connectable tangent planes. Finally, the interpolation of distance value is performed using an upsampling filter on each region consisting of superpixels that have smooth-connectable tangent planes. The clustering of tangent planes using ray-tracing calculation achieves efficient and noise-robust smooth surface segmentation, and a high completion rate is achieved in noisy situations without decreasing the interpolation accuracy. In additional, ray-tracing calculation for tangent planes makes it possible to estimate distance information of surfaces behind occlusions. In experiments on images from the Middlebury stereo datasets with added simulated Gaussian measurement noise, our method interpolates each image at a high completion rate, and achieves low error when compared with existing techniques.
In this paper, we discuss the development of an inspection system for a gloss-coated surface using patterned illumination. The convex defect on a gloss-coated surface is caused by top-coating paint on a primary coating with minute particles such as dust remaining. Since the convex defect is transparent, it is difficult to observe it in conventional illumination. Thus, we developed an optical system with patterned illumination and an inspection system using imaging technology with a phase-shifting method given the behavior of specular reflection on a gloss-coated surface. The inspected surface is illuminated with the patterned illumination by shifting the phase of a stripe pattern, and a camera takes multiple images of the specular reflection. By calculating the amplitude of the luminance modulation according to a phase-shifting method, the amplitude image can be obtained from the multiple images. The amplitude image means the distribution of the reflectance. The scratch and dirt as well as small convex defects on a gloss surface can be observed in the amplitude image. This inspection system can make an image of the shape and specular reflectance on a gloss surface and allows inspection of gloss coating, which was difficult in the conventional method.
The present method of getting information of road signs has problems about working hours and expense because it is manual labor. We are developing a system about automatically recognizing of road signs which consist of characters, e.g. signs of intersection's name using on-vehicle camera. In order to increase a recognition rate, we adopted Learning Based Super-Resolution (LBSR). LBSR's resolution enhancement depends on a similarity degree of filmed road sign image and low-frequency components of learning image. In this paper, we propose a reasonable method of generating low-frequency components of learning image for road sign images. The experimental result shows that the proposed method slightly improved the recognition rate.
This paper introduces a high-speed 3-D object recognition method using a novel feature description. Features proposed in this study consist of three values. One is the Difference of Normals (DoN) feature value that has been proposed by Ioannou. The other two represent information about curvature. These features are named Combination of Curvatures and Difference of Normals (CCDoN) Features. These features are used for recognition of position and pose of multiple objects that are stacked randomly. Because they are low-dimensional, high-speed matching can be achieved. Moreover, high-speed and reliable matching is achieved by using only effective features selected on the basis of their estimated distinctiveness. Experimental results using real datasets have demonstrated that the processing time is about 81 times faster than that of the conventional Spin Image method. Furthermore, the proposed method achieves a 98.2% recognition rate, which is 46.6% higher than that of the Spin Image method.
In this paper we address the problem of image classification by embedding the spatial information into the local descriptor. In our method, we directly concatenate (x, y) coordinates of an image into the original feature vector. This simple idea can perform well in the object category classification even though the feature vector size is almost the same as the conventional approach. Results are reported for classification of the Caltech-101 dataset and our methods are found to produce consistently better results compared with traditional Bag-of-Features approaches in all experiments.
In the visual inspection automated by the image processing, the importance of the machine vision lighting technology have recently drawn attention among the users of the visual inspection device. The machine vision lighting technology provides methods of the selection and the adjustment of illuminations for the visual inspection appropriately. However, it is difficult for the users to study the machine vision lighting technology because it needs the knowledge and the experience. Moreover, a defective product has several defective features appeared or disappeared with the angle and the color. This paper proposes the system to adjust parameters of the illumination for the visual inspection. A non-defective product is captured by this system with modifying four parameters : light quantity of red, green and blue LEDs, and the angle of LEDs. Further, the appropriate parameters of the illumination are decided by the spectrum analysis used for the image quality evaluation. In the step of evaluation, three type of defective/non-defective products are used to evaluation. Parameters of the illumination are decided by our system using the non-defective product. All of appearance of defective-features are confirmed by the visual observation and binarizing images in all type of products using this evaluation.
Ball trajectory is an important performance indicator in ball game. However, Measuring 3D trajectories is often unusable because synchronized cameras and control points inside the playing field are generally required. A method that reconstructs 3D trajectory of ball with unsynchronized cameras is proposed in this paper. Proposed method consists of ball detection, camera calibration, and trajectory reconstruction. At first, ball candidates are detected based on the appearance. For the camera calibration, highly probable balls are extracted from the candidates based on the motion. Imaged ball trajectories are then reconstructed. Next, fundamental matrix is computed from the corresponding points, which are estimated with the ball trajectories, and roughly estimated temporal offset, which is estimated manually. The estimated fundamental matrix is assumed to be inaccurate due to the inaccuracy of the temporal offset. The key feature of this method is to accurately estimate temporal offset and fundamental matrix by optimizing them iteratively based on the error of fundamental matrix. After the calibration, balls are extracted robustly with geometrical and temporal relationship between cameras. A ball trajectory is finally reconstructed as connected trajectories which separated at collisions. It is experimentally demonstrated that the proposed method can calibrate cameras precisely and reconstruct accurate ball trajectories.
In this paper, we present a robust pose estimation method of a human body hidden with a cover. Our goal is an automatically monitoring system of sleeping human using noncontact and noninvasive sensors for elderly care. We propose a new method for robust human pose estimation from a single depth image using human body shape model constructed by normal vector information. This shape model is able to represent shapes of rough body, and is effective in robust pose estimation for a person who placed futon and blanket. In our method, first, head position is detected from a depth image using SVM. Then, body region is detected by comparing human body shape model. Head position is used as initial position of body region detecting. Body region is composed of many small rectangles. Next, body region is divided into three body parts by distance between parts. Finally, each part pose is determined by a linear estimation using point clouds of right and left of the rectangle in each body region.
This paper introduce a 3D position and pose recognition method that can recognize various shape objects even if appearance of the object model has planner shape. In order to handle these case, the proposed method automatically selects suitable matching strategy for shape aspects of each segment, after applying the segmentation to an input range data. If the segment has a lot of local shape features, such as 3D keypoints, the local feature based matching is applied. On the other hand, if the segment has less local feature, the Global Reference Frame (GRF) which represents 3D dominant orientation of the segment is generated, and it is used for recognition. The GRF consists of two independent vectors. One is the dominant surface normal vector of the segment, the other is the dominant orientation vector of projected segment's range data onto tangent plane. By calculating the differential orientation of pair of the GRF, rigid transformation that align two segments will be calculated. Experiments have confirmed that the proposed method increases the recognition success rate from 84.1% to 94.7%, in comparison with the state-of-the-art method.
Bullet identification is the difficult work that only specialists can do. It requires a high level of skill and takes a long time. So we are trying to assist appraising workers by using an image processing technology. In this paper, we propose a method to express the similarity between two bullets numerically and to estimate whether two bullets were fired by the same firefream.
Land cover maps are made from satelite sensored data, utilizing the estimated reflection model on the earth. Since the accuracy of the land cover maps is about 60%, some calibration is needed. Science DCP is a project to be used for scientific validation data that visited the point where the terrestrial intersections of integer degrees of latitude and longitude. The purpose of this paper is the automation of land cover classification. We develop data of landscape images of 4211 points on global. We experimented image recognition using SVM and BoK. In this case, we found that subclass classification by k-means algorithm is effective.
In this paper, we discuss a new visual tracking algorithm based on resampling images obtained from virtual view-points. An area depth sensor can capture Point-Cloud-Data of real 3D-scene. The virtual view-points can be located everywhere in a reconstructed scene space. If the virtual view-points are located on a motion vector of a tracking target object, the target object appears at rest in the resampling 2D-image. Therefore, it is easy to track a motionless object. We have implemented above-mentioned concept. Experiment results indicate that our approach is reasonable.
In this paper, we propose a new visual inspection method that can analyze the exterior surface of rubber tires by means of the light stripe projection. Image sensing of the tire surface was implemented by setting the tire on the rotating table so that the normal pattern of the tire surface could be suppressed. The proposed method detects the exterior thin defect after removing three kinds of uneven circularities due to the real shape of tire, the eccentricity of rotating table and the inconsistency between two. Thus, in this research, we realized a visual inspection method by transforming three-dimensional shape of tire surface into two-dimensional image information through the analysis of light stripe images.
Image morphing method can generate new emphasized images smoothly from input images by extrapolation. It is one of the useful methods for making an impressive movie. However, the extrapolation of image morphing has a problem that the output images often have dynamic range higher than its input images. Generally, at a device for the input images having low dynamic range, we cannot display the output images having higher dynamic range. In this paper, we propose a dynamic range compression method for the images generated by extrapolation of image morphing. This method compresses the dynamic range into it having the input images, while it preserves image visibility and smooth variation from the input images. This method achieves a reproduction of the emphasized image by extrapolation of image morphing, without high dynamic range device. In the experiment, we confirmed the visibility quantitatively and smooth variation from the input images qualitatively.
As the recent improvements in appearance rendering capabilities of computer graphics(CG), CG technologies are widely applied to various field. In this research, we intend to generate high quality CG of woven cloth, and focus attention on fluorescence and transparent characteristics. First of all, we measure excitation spectra, fluorescence spectra and Bidirectional Transmittance Distribution Function(BTDF). Secondly, we combine these measurements and propose a new BTDF approximation model with considering fluorescence color and intensity(Multi-band BTDF model). Finally, we generate lace curtain CG images using this Multi-band BTDF model. The images suggested that fluorescence characteristics are important factor on expression of woven cloth using computer graphics.