Takaaki Kaiga

Adaptive Keypose Extraction from Motion Capture Data

In this paper, we present a novel method to extract keyposes from motion-capture data streams. It adaptively extracts keyposes in response to the motion characteristics of a given data stream. We adopt an approach to detect local minima in the temporal variation of motion speed. In the developed algorithm, the intensity of each local minimum is first evaluated by using a set of signals; it is obtained by applying a set of low-pass filters to a one-dimensional motion-speed data stream. The cut-off frequencies of the filters are distributed over a wide frequency range. By adding up the speed-descent values of each local minimum over all the signals, we exhaustively obtain the information on its intensity provided at all the time-scale levels covered by a given data stream. Then, the obtained intensity values are categorized by a clustering algorithm; the local minima categorized as those of little significance are deleted and the remaining ones are fixed as those giving keyposes. Experimental results showed that the present method provided results comparable to the best of those given by the methods previously proposed. This was achieved without readjusting the values of parameters used in the algorithm. Readjustment was indispensable for the other methods to obtain good results.

Development of a high precision hand motion capture system and an auto calibration method for a hand skeleton model

Motor cortex that controls the movements of the human body is divided functionally into each control region. In the motor cortex, the area for the hand is almost the same as the total area for arm, torso and lower body. This physiological fact indicates that human hand movements require very complicated control. As a result, our hand can perform high precision movements as an actuator. Motion capture (MoCap) technique that can digitize a position and a posture as a function of time is widely used in order to create animation and CG. It is very difficult to measure all hand movements because one hand has twenty-seven bones and nineteen joints. Therefore, it has been impossible to record the finger movements of a sports player that are high in speed and in accuracy.

A hybrid approach to keyframe extraction from motion capture data using curve simplification and principal component analysis

In this paper, we propose a novel method to extract keyframes from motion capture data. A hybrid approach, which combines a curve-simplification algorithm with an initialization procedure including principal component analysis, is adopted. The developed method automatically extracts an appropriate number of keyframes at high speed without performance degradation. Experimental results prove the effectiveness of the present method.

Indexing of Motion Capture Data Using Feature Vectors Derived from Posture Variation

Recently several large-scale databases of motion-capture data streams have been constructed. We present a novel method to index motion-capture data streams in such databases. We pay attention to posture variation; the impression of the visual aspect of the whole body is regarded as important. The spatial distribution of body segments is statistically summarized as a feature vector having only 12 dimensions. The experimental results showed that the feature vector we introduced provided properties comparable to those of the methods previously proposed, even though its dimensionality is extremely low.

Adaptation of Grouping Structure Analysis in GTTM to Hierarchical Segmentation of Dance Motion

In this paper, the authors propose the adaptation of the rules used in the grouping structure analysis in Lerdahl and Jackendoffs “A Generative Theory of Tonal Music (GTTM)” to dance motion analysis. The success of the adaptation realizes the segmentation of dance motion in a hierarchical fashion. The analysis method obtained by the trial of the above adaptation consists of the following procedures. A motion-capture data stream of a dance is first divided into a sequence of events by piecewise linear regression. The hierarchical structure of groups each of which consists of a sequence of the events is then extracted by applying the grouping rules adapted to dance motion analysis. The above method is applied to motion-data streams acquired by motion capture systems. The obtained results indicate the following advantages: (1) The structure of hierarchical segmentation is precisely extracted in response to the characteristic of an analyzed dance. (2) The extraction of the hierarchical segmentation provides the possibility of the development of a technique distinguishing the oversegmentation from regular boundaries. (3) The possibility of utilizing the information of hierarchical segmentation for the comparison of dance performances is suggested.

Extraction of Motion Characteristics in Dances by Statistical Analysis of Joint Motions

In this paper, the authors attempt to develop a technique for the analysis of the motions of dances having no stylized motion structure, focusing on joint motions. The variance-covariance matrix given by the statistical analysis of the time-series data of joint motions is selected for the evaluation index characterizing dance motions. The application of the derived evaluation index to the representation of dissimilarity between dances is shown to be effective when the whole commonness appearing in both the dances compared should be considered. It is also confirmed that the application of multidimensional scaling (MDS) with the orthogonal rotation of coordinate axes is effective to extract the distribution feature of a database of dances. The evaluation items characterizing all the dances belonging to the database are automatically extracted by the analysis of correlation between the coordinate axes given by MDS and the elements of the variance-covariance matrix.

Multi-level segmentation of dance motion by piecewise regression

It has been recognized that a technique to divide a raw motion-capture data stream of a dance into segments on the time axis is needed [Sonoda 2008]. In particular, the extraction of the higher-level information such as the hierarchical segmentation-structure is a subject of growing interest at the present time. In this study, the authors attempt to develop a method to segment dance motion in a multi-level style, namely in a hierarchical fashion.

Development of a Visualization Method for Motion-characteristic Distribution of Japanese Folk Dances - A Case Study of the Bon Odori Dance

This study proposes a method to systematically visualize the motion-characteristic distribution of Japanese folk dances passed down in a certain area. This is accomplished by adopting an approach that involves analyzing motion-capture data collected from the dances. The visualization process in the proposed method consists of three stages. The first stage is the modeling of the relationship among motion-capture data, folk dances, and the settlements in which folk dances have been passed down. This relationship is modeled as a hierarchical-structure model. The second stage is the extraction of motion characteristics from motion-capture data streams. The motion characteristics of each data stream are summarized as a fourteen-dimensional feature vector. The third stage is the visualization of the motion-characteristic distribution of the dances investigated. Each of the dances is mapped on a two-dimensional scatter plot in accordance with the feature quantities obtained in the second stage. Information on the hierarchicalstructure model constructed in the first stage is also displayed. The analysis results for the distribution of Bon Odori dances showed that the proposed method could have almost completely visualized the motion-characteristic distribution of sample folk dances, while also demonstrating consistency with the knowledge of the dances acquired in the previous studies.

Low-dimensional Feature Vector Extraction from Motion Capture Data by Phase Plane Analysis

This paper proposes a method to obtain a low-dimensional feature vector appropriately representing the characteristics of a given motion-capture data stream. The feature vector is derived based on the concept of phase plane analysis. A set of phase plane trajectories are obtained from the temporal variation of the state variables representing the body-segment arrangement. The information on six motion-characteristic properties is extracted from the shapes of the trajectories, and used as the components of a six-dimensional feature vector. The experimental results showed the effectiveness and limitation of the proposed method.

A motion style estimator for lost folk dances in Akita prefecture, Japan

In Akita Prefecture located in the northeast region of Japan, many attractive folk dances have been passed down. However, several dances were lost due to the falling birthrate and depopulation of rural areas. This study proposes a method to obtain the information contributing to the restoration of the lost folk dances. Specifically, we estimate the motion style of lost Bon Odori dances (Bon Odori: a type of Japanese folk dance), using the investigation data of folk customs and the motion capture data of extant dances.