This paper presents real-time human motion analysis based on silhouette contour analysis. Our purpose is to develop a computer-vision-based human motion analysis system, which can be applied to human-machine interaction via human gestures. Since vision-based human motion capturing systems are essentially unstable and can only acquire partial information because of self-occlusion, we have introduced a robust pose estimation strategy, which can estimate human postures from limited perceptual cues such as positions of a head, hands and feet. In this paper, we outline a real-time and on-line human motion capture system and demonstrate a simple interaction system based on the motion capture system.
We have developed a landscape simulation system and have promoted its practical use in various area. Our system has been utilized in some fields and has proved to be effective for visual communication between designers and users. In the system, orthophotos calculated from aerial photography are texture-mapped to high-precision 3-dimensional terrain data to move their viewpoint smoothly by real-time rendering device. In this paper, a high-precision and interactive landscape simulation system developed by us is reported, and its practical use is overviewed with its possible development in future.
Until now, many thing 3-D image measurement techniques have been proposed. The measurement which projects a projection pattern is also one of them. However, if measurement environment and a measurement object change, the accuracy of measurement will become bad or the measurement will become impossible in the case of the fixed measurement parameter is used. In order to solve the above-mentioned problem a technique is proposed to control the size and the intensity distribution of the projection pattern. If the optimum projection pattern can be projected according to a measurement scene, more active 3-D image measurement will be attained. This paper describes mainly the technique of the intensity control of a projection pattern automatically.
Fisher's discreminant analysis achieved the robust recognition independent of the lighting condition. This technique, however, uses the pixel of the image as an element for the distinction calculation, so calculation load becomes huge. Moreover, the number of registration images becomes enormous depending on the range of a change in lighting. This study then proposes a method to represent a face by 12x12 mosaics instead of picture elements, and the mosaic elements, where changes in the brightness become large by the shadow as change in the lighting angle becomes large, are removed from the calculation elements. Moreover, the number of dictionary registration reduced by complementing the dictionary of a middle angle which are sandwiched by adjustment angles. As a result, this method realizes an average of 94.6% of the high rate of recognition over the change of four 45 degrees directions.
The importance of picture retrieval is increasing by reason of spread of multimedia content, these days. Picture retrieval with rough sketch is one of efficient methods, when we don't express our target picture uniquely with some words or when our target picture was tagged inappropriately. We had problems, though, that many candidates were shown by retrieval system to our rough sketch. We propose to solve the problem that we add object's motion information to our rough sketch, when we retrieve our target picture from moving pictures. Supposing that we use MPEG-7(ISO/IEC 15938)as metadata, we consider that how to interpret user's rough sketch and motion information.
A similarity search method is presented for videos based on their segmentation by clustering. Videos are partitioned into shots by clustering of the set of frames based on color histograms. The performance of this segmentation is improved by addition of temporal information to color data. Videos are represented by mixture distributions and the distance between videos is evaluated by the works needed for transforming one distribution to another one. This distance is used for similarity search of videos. It is also shown for this similarity search that its performance is raised by incorporating temporal information. Finally a filtering method based on an inequality for the distance is devised for speeding up the similarity search.
A method is presented for similarity search of videos by using temporal correlogram. Each video is transformed to a sequence of symbols by vector quantization. The symbols are obtained by clustering of a sample set of frames based on quadratic from distance of color histograms. Earch frame of videos is coded to the symbol of its nearest neighbor. This coding procedure is accelerated by filtering based on the triangle inequality. Features of sequences of symbols are represented by temporal correlogram. Videos are retrieved on the basis of distance of the temploral correlograms. Computational time of the retrieval is reduced by filtering based on an inequality of the distance of the temporal correlograms.