Search

The purpose of this paper is to propose a new approach to obstacles detection using a single camera mounted on a vehicle when the vehicle is backing or turning round at an intersection at a low speed. Using equations among feature point locations and optical flows in geometrically transferred top-view images, ground-movement information can be estimated. We compensate for the ground-movement between consecutive top-view images using the estimated ground-movement information and compute the difference image using the compensated previous top-view image and the current top-view image. Finally, obstacle regions on top-view image can be easily extracted using the difference image. The proposed approach is demonstrated to be effective by evaluation images captured by an in-vehicle camera

We describe a practical and precise calibration method for generating a top view image that is transformed so that a planar object such as the road can be observed from a direction perpendicular to its surface. The geometric relation between the input and output images is described by a 3×3 homography matrix. Conventional methods use large planar calibration patterns to achieve precise transformations. The proposed method uses much smaller element patterns that are placed in arbitrary positions within the view of the camera. One of the patterns is used to obtain an initial homography. Then, the information from all of the patterns is used by a non-linear optimization scheme to reach a global optimum homography. The experiment done to evaluate the method showed that the precision of the proposed method is comparable to that of the conventional method where a large calibration pattern is used, making it more practical for automotive applications.

We studied the relationship between purchasing behavior and physiological measurement using an eye camera, including the correlation of physiological measurement based on an analysis of eye movement data at the behavioral level (close to intention) using conjoint analysis and with the attitude level using the analytic hierarchy process. Results showed poor correlation between data on eye movement, i.e., physiological reaction to stimulus, and the attitude level but a fairly good correlation between data on eye movement and the behavioral level.

We propose a high speed iris detection method for man-machine interface. In order to communicate between humans and computers comfortably, it is required that high speed processing and adapting the changes of the face direction and the individual variation of the face. Our iris detection method consists of a face detection function, a facial parts extraction function and an iris detection function. The face detection function adapts to individual variation or a light condition by using standard skin color method which decides the standard skin color by a color histogram every frame. The facial parts extraction function adapts to changes of face direction and the individual variation in order to by using the hybrid template matching which consists the four directional features field and the color different features. The iris detection function performs high speed and stable detection by segmentation of iris region using color information and the Hough transform. We conducted the comparison experiment with our past method which uses the least approximation method. By this result, processing time of our method realized equivalent to our past method. In addition, accuracy rate of iris detection with our method improved.

To track and recognize the movements of multiple people using multiple cameras, each person needs to be segmented and identified in the image of each camera We describe a method that tracks multiple people and identifies their body parts in multiple camera images Estimating the positions of people and identifying the multiple cameras are mainly based on the silhouette method combining the background and frame subtraction Our method was improved by using the existence probability map on noise problems The existence probability map consists of the position probability and height cumulus maps. A gesture event is detected by comparing the human and motion blocks that are extracted from the voting space using the silhouette method We demonstrate experimental results and the validity of our method

We propose the use of an adaptive discriminant space constructed from facial images of candidate faces to optimally distinguish the faces of certain candidates. Firstly, our method to select the set of candidates uses a hierarchical combination of the four-directional features for face-direction estimation and recognition based on linear discriminant analysis. Next, the exact directions of the faces are estimated from the respective discriminant spaces for face-direction estimation. Finally, our method creates a new discriminant space from the facial images of candidates that were selected on the basis of the results of the procedure for face-direction estimation. This adaptive discriminant space is used in opimally selecting from among the candidates. Experiments showed that our method improved the accuracy in recognition by 1.6% to 99.2% by hierarchical face recognition of 105, 000 images from 150 subjects facing in 35 different directions.

We propose a method to detect face regions by using the standard skin color which is taken from an image in a color histogram. Extraction of the face regions is based on an area in motion and the deviation from the standard skin color obtained by discriminant analysis. The standard skin color is updated from information obtained from the detected face region. To judge a face, we use templates composed of directional features and color features to detect eyes and mouth. In the face detection experiment, our algorithm could detect a face region by using motion of a subject's head. In addition, the updated standard skin color was correct. In the facial parts detection experiment, by updating the templates, our method worked better than the method that used only directional features. Consequently, our algorithm can extract the facial features even when a person tilts his or her head. With a SGI/O 2 workstation, it is possible to achieve a speed of approximately 0.25 sec/frame.

To clarify the human mechanism that controls the relation between information processing and eye movements, we propose a fixation masking method that detects fixations and masks only those fixations. This paper describes how we examined the time required to obtain information on Japanese characters as well as the quantity of information acquired at each fixation point. Subjects were instructed to recognize incorrect characters from text they read. The results indicated that at least 67 ms were required to obtain visual information after viewing the stimulus. However, the masking affected the character recognition process in brain. From the distribution of fixation duration, approximately 130 ms were required to invoke the saccade once the subject had perceived the mask. Our results were different from Rayner's results. We assume that more complex information processing is required to read Japanese than to read English.

The purpose of this paper is to propose a new approach to detect obstacles using a single camera mounted on a vehicle when the vehicle is backing or turning round at an intersection at a low speed. Using equations among feature point locations and optical flows in geometrically converted top-view images, ground-movement information can be estimated. We compensate for the ground movement between consecutive top-view images using the estimated ground-movement information and compute the difference image using the compensated previous top-view image and current top-view image. Finally, obstacle regions in top-view images can be easily extracted using the difference image. The proposed approach is demonstrated to be effective by evaluation images captured by an in-vehicle camera.

This paper describes a practical calibration method to generate a top view image from an image of a rear view camera for automobile. The geometric relation between the input and output images is described by a 3 x 3 homography matrix. Conventional methods have to use a large planar pattern to achieve precise calibration. The proposed method uses much smaller patterns, and enables to place those patterns on arbitrary positions within the view of the camera. The experimental results show that the method has equivalent calibration precision with much easier setting than the conventional methods.

Recently, gender estimation from face images has been studied for frontal facial images. However, it is difficult to obtain such facial images constantly in the case of application systems for security, surveillance and marketing research. In order to build such systems, a method is required to estimate gender from the image of various facial poses. In this paper, three different classifiers are compared in appearance-based gender estimation, which use four directional features (FDF). The classifiers are linear discriminant analysis (LDA), Support Vector Machines (SVMs) and Sparse Network of Winnows (SNoW). Face images used for experiments were obtained from 35 viewpoints. The direction of viewpoints varied ±45 degrees horizontally, ±30 degrees vertically at 15 degree intervals respectively. Although LDA showed the best performance for frontal facial images, SVM with Gaussian kernel was found the best performance (86.0%) for the facial images of 35 viewpoints. It is considered that SVM with Gaussian kernel is robust to changes in viewpoint when estimating gender from these results. Furthermore, the estimation rate was quite close to the average estimation rate at 35 viewpoints respectively. It is supposed that the methods are reasonable to estimate gender within the range of experimented viewpoints by learning face images from multiple directions by one class.

A facial parts detection technique that is robust for face poses is described. The initial probability is obtained by template matching of the four directional features. Positions of facial parts are detected by relaxation matching using the spring connection of each facial part. The relaxation method has a problem that requires much computation time. We propose a high-speed relaxation technique for facial parts detection. Experiments show that facial parts are detectable to the multi-directional face image database.

We propose a multi-camera system that can track multiple human faces and hands as well as focus on them for face and hand-sign recognition. Our current system consists of four video cameras. Two cameras are fixed and used as a stereo camera to estimate position. The stereo camera detects faces and hands using a standard skin color method we proposed. The distances of targets are then estimated. To track multiple targets, we evaluated position and size of targets in consecutive frames. The other two cameras perform target tracking. Our system selects a target for recognition by using size and motion information sequentially. If size of the selected target is too small for recognition, tracking cameras acquire its zoomed image. Using our system, we experimented on multiple target tracking.

This paper proposes a method for estimating continuous face directions with the discrete face direction feature classes. The feature classes are composed by the linear discriminant analysis of the "four directional features" that are extracted from some people's faces looking in specified directions. Estimation by a single camera gives ambiguous results for non-frontal view of faces. Especially, it is impossible from the back of the face. However, integrating the information from the multiple cameras allows more precise face direction estimation. Experiments show that our proposed method can achieve a mean error of at most 8° and a variance of 22.4 for all face directions.