実世界知能技術分野

Theme and Laboratory	Outline
Hand Motion Estimation based on a Generic Model Multi-Modal Functions Sanyo Laboratory http://www.rwcp.or.jp/activities/achievements/MM/sanyo/index-j.html	A method for manipulating 3-D objects by tracking the rigid motions of the human hand is presented. The essential points are as follows. 1)This method first detects only points on the finger boundaries in agray-scale image taken by a monocular camera by using shape primitives based on a ''generic model'' concept. 2)The method then determines the tracking points and detects corresponding points from frame to frame using local edge features in the image sequences such as a tangent vector at each point in the first frame and a change of intensity along this vector. 3)Next, parameters of the hand motion are calculated using the coordinates of the corresponding point pairs in each image under the constraint of planar surface motion. Finally, the 3-D model of the object can be manipulated on a PC by changing its position and orientation with calculated motion parameters.
Which areas of the brain are involved when reading Japanese characters? Multi-Modal Functions NTT Laboratory http://www.brl.ntt.co.jp/cs/rwcp/index-j.html	In the multimedia information era, various kinds of visual and auditory information are provided. These various kinds of information are in fact processed in our brain when we understand the information. Aiming at better method to provide information that is easy to understand, we started to investigate the brain activity associated with information processing.Using the neuroimaging method, we recorded the human brain activity when the subjects were reading Japanese characters to identify which areas of the brain were involved in this reading process. The brain emits very weak magnetic fields when it is activated. Analyzing the brain's magnetic fields while the subject read kanji characters, we localized the active regions in the occipital, parietal, and temporal areas of the brain. According to the neuropsychological studies, the occipitotemporal area is involved in reading kanji characters while the occipitoparietal area is activated in reading kana characters. Our results, however, showed that both areas were activated in reading kanji characters.
Semantic information Retrieval from Tagged Text Corpus Information-Based Mitsubishi http://www.rwcp.or.jp/activities/achievements/IB/mitsubishi/achieve_j.html	In response to the query made to the system "Are there politicians who want to promote political reform as an economic measure?" we can retrieve from the database of newspapers the response "Deputy Prime Minister/Minister of Foreign Affairs, Hata went so far as to say, 'political reform is the best economic measure to boost demand.'" Conventional retrieval systems only allow us to retrieve a list of documents (typically many in number) that include the specified keywords. However, the present system allows us to directly retrieve the sentence that is a response to the query and the context in which it appears, in an interactive manner. If the object to be retrieved is plain text, the implementation of semantic information retrieval becomes very difficult. Semantic information retrieval, however, becomes possible without waiting for the solution of various subjects (ellopsis, anaphora, disambiguation, etc.) in natural language processing technology if we attach tags to the text of the object to be retrieved beforehand. Although the principle of the present system is still being verified, we are planning to complete a prototype in two years through further research and development.