Abstract:

Kinematic analysis of human motion and behavior has intensively performed in order to obtain clues for enhancing sport performance. In kinematic analysis through conventional biomechanical approaches, parts of interest on a moving object such as human joint centers are detected and tracked regardless of with- or without-markers, and then motion characteristics are extracted from the resultant trajectories. Motion capture technologies enriching such kinematic analysis have been, in most cases, used in experimental context. Existing approaches employed in practical context have more or less difficulties in feeding back the analysis results for athletes and coaches timely according to their requirements.

In recent years, on the other hand, video-based kinematic analysis using computer vision and pattern recognition techniques has attracted a great deal of attentions by virtue of the applicability in practical context. Several attempts to develop a method for extracting the characteristics of the full-body human motions in an automated manner have been made. However, many of ordinary approaches require specific and tedious steps including region segmentation of interest, in which the error tends to affect final result. Kobayashi and Otsu (2006, 2009) have proposed a scheme of adaptive vision system employing cubic higher-order local auto-correlation (CHLAC) motion features and then have successfully applied to the wide range of motion recognition tasks with their colleagues.

CHLAC features characterize object motion, reflecting kinematical measures such as motion orientation, velocity and acceleration. In addition, CHLAC requires neither prior knowledge, heuristics about objects nor time-consuming computational cost. Such the properties of CHLAC are vital for motion analysis in sport. CHLAC motion features can be also widely applied to motion recognition tasks as well as motion analysis tasks.

This paper presents a method for estimating dominant motion orientation and magnitude of objects, applying directionally-grouped cubic higher-order local auto-correlation (DgCHLAC) motion features. The DgCHLAC motion features are extracted based on the CHLAC feature components classified into predefined direction groups. Dominant motion orientation and magnitude of objects can be simply estimated by identifying the dominant DgCHLAC feature component with respect to magnitude.

The experiment was performed to examine whether the method was useful for extracting motion characteristics, and the resultant motion indicators were also of practical use for performance analysis in sport. The experimental results can exhibit the applicability of the proposed method to the motion orientation estimation. The proposed method can be utilized not only for more refined and higher-level motion analysis such as characterization of each motion and also skill, but also efficient video feedback and browsing.

Keywords:

Video-based motion analysis, automatic motion feature extraction, effective coaching.