Search

This paper propose a method of writer verification basedon on-line features extracted from a process of drawing afigure, and evaluate its performance on handdrawn figuresby genuine writers and pretenders. In the proposed method,a person in registration writes only the reference figure,which is composed of as many parts as possible and theperson can be verified with a arbitrary part of the referencefigure. The verification is based on a spotting methodcalled CDP (Continuous Dynamic Programming). Fromsome requirements in authentication system, the proposedmethod is tested through some experiments using figures.The error rates in the experiment were different dependingon conditions such as kind of on-line features, minimumnumber of strokes comprised in acceptable test sequence asinput, figure for determination of threshold and the genuinewriter. The result of the experiment shows that the proposedmethod is effective.

The purpose of the paper is to propose an algorithm for pseudo-phonemic labelling of each frame feature in a continuous speech. The algorithm is consisting of the two key concepts; 1) Utilization of speech feature composed of spectrum and vector field feature of spectrum based on Vector Field Model of Speech for realizing a system which works speaker-independently, and 2) Application of a new algorithm for accumulating pseudo-phonemic evidence of each frame based on both Continuous Dynamic Programming (CDP) and Partial Matching Method.

A new model called Vector Field Model is proposed for providing new algorithms of both segmentation and feature extraction in order to recognize phonemic units in continuous speech spoken by many speakers. The original vector field is obtained by differentiating a time-frequency pattern (the output of band-pass filters). In order to extract steady , increasing transient or decreasing transient feature of the point on the time-frequency pattern, three auxiliary vector fields are created by characterizing coherent orientations of vectors. The crowded vectors in an arbitary auxiliary vector field produce a pseudo-phonemic segment. Recognition of /VCV/ is carried out by applying so-called Continuous Dynamic Programming to a segment sequence pattern.

We propose a method to self-organize a user model for realizing a text search system. The method self-organizes a model from phoneme sequences and their texts using 6 basic rules. An inference method associated with the self-organized model searches texts corresponding to inputted phoneme sequences. Experiments using 50 voice waveforms of 10 speakers show that the self-organized model is useful to realize a text search system and is robust for noisy speech utterances.

We develop a technique for robustly positioning and guiding mobile robots in the real world, such as offices and homes, by using topological maps consisting of image sequences capturing the environment. This technique uses reference interval-free continuous dynamic programming (DP) which allows positioning by spotting, but it restricts the direction of mobile robot navigation to the same one as the map creation. We propose to use the nonmonotonic continuous DP for gesture recognition, to allow the positioning of a robot within the same framework even when it navigates backwards or stops. We also propose the use of average values of local features in each section generated by horizontally dividing the omnidirectional view as rotation-invariant features to allow global positioning of a robot regardless of its direction. In addition, we demonstrate the usefulness of this technique through evaluation experiments using mobile robots.

We propose a method to recognize human gestures using both motion images and text. The method uses two kinds of network models obtained from two kinds of sequences in a style of self-organization. We can extract both so-called common and singular parts of a gesture by analyzing the topology of the network of gesture motion images. If the order of movements in a gesture motion image matches with that of words included in a corresponding sentence, then we can also extract both so-called common and singular parts of the network model of test. The proposed method for recognizing gestures uses morphological features between two networks made of gesture motion images and word sequences. We show the usefulness of the method through an experiment using a database composed of pairs of gesture motion images and text.

We propose a method to recognize human speech without the restriction that the speaker should be close to the microphone for the multiple modal interface. This method has the following features: 1) to control the microphone directivity for improvement of S/N ratio by using the speaker's position estimated by the result of gesture recognition, 2) to compose the impulse response of the transfer characteristic between speakers and microphones into reference data of the recognition system. We describe the active microphone method in detail, and show its effectivity through several experiments. Then, we discuss the results.

A spotting algorithm to recognize the meanings of human gestures from motion images is proposed. Feature vectors are extracted from temporal edge image. And input image sequences are matched with models by spotting algorithm. The spotting algorithm removes the need for temporal segmentation of gesture duration and introduces frame-wise recognition. In this paper, a new spotting method called nonmonotonic continuous DP is proposed for spotting gestures and their variations such as reverse, partial and stop motions. We showed the effectiveness of the new spotting method.

This paper proposes two new methods of how to carry out automatic topic spotting in continuous speech and how to describe its summary in the form of the slot expression. The topic spotting works independently of both speakers and topics. The method of topic spotting is based on a dialogue model to segment a topic using only a surface feature of continuous speech. The spotted speech are selected by the co-occurrence relations and are classified by the word thesaurus. We evaluated our methods using the speech of the broadcasting news and the spoken dialogue speech. The result shows that detection quality was over 80% on the broadcasting news.

A retrieval method, called running time interval clustering (Rutic), is proposed. The method detects time sequence intervals similar to the query time sequence in a vast database of time sequences such as video and audio data. Conventional methods, such as RIFCDP, handled queries of any interval lengths, but they require a relatively high computational burden and are not suitable for real-time retrieval from a large database. The Rutic method allows retrieval results to be output for each input frame of the time sequence queries, thus enabling retrieval from a database without segmentation. The computational burden of the Rutic method is so small that it enables the so-called real-time spotting retrieval. This report describes the algorithm of the Rutic method, and verifies its validity through comparison with other methods based on image search experiments.

In order to obtain flexibility of recognition system of human motion behaviors, omni-directional vision sensors are used for both tracking humans and getting multiple. views of different directions that allows human to move with natural behavior in large area. First, a vision-based aspect model is proposed which realize the recognition independent of the human-direction. It contains feature vector sequences that were taken from all direction around the human. Next, the authors' appearance-based feature extraction method is demonstrated to be useful for whole body motion recognition. A new technique, spatio-temporal continuous dynamic programming, allows a human to rotate even when he is doing a motion behavior.