DIGITAL LIBRARY
EFFICIENT COLLECTION OF SIMILAR SPORT SCENES OF INTEREST FOR EFFECTIVE COACHING
Biwako Seikei Sport College (JAPAN)
About this paper:
Appears in: INTED2015 Proceedings
Publication year: 2015
Pages: 6103-6107
ISBN: 978-84-606-5763-7
ISSN: 2340-1079
Conference name: 9th International Technology, Education and Development Conference
Dates: 2-4 March, 2015
Location: Madrid, Spain
Abstract:
In sport coaching and physical education (PE), utilizing videos effectively are vital for encouraging and facilitating athlete’ and student’s performance enhancement activities, including motor learning. Sport videos can also promote athlete-coach and student-teacher interaction, and provide much information on sport performance.

Analysis and evaluation of motions and tactics in sport is often conducted on specific video scenes from a similar viewpoint. Detecting and segmenting similar scenes of interest in such sport videos is a fundamental procedure for utilizing the video contents to facilitate effective coaching. However, existing routines for video-based performance analysis in sport and physical education (PE) involve heavy and cumbersome labor in both the data processing and handling, such as seeking and extracting scenes of interest in large number of videos. Consequently, conventional video feedback systems have difficulties in providing the video contents for coaches and PE teachers timely according to their requirements. Thus, it is a crucial issue to detect and segment specific and similar scenes of interest in an automated manner for efficient video observation and further analysis.

Many scene detection methods have been developed and applied to sport videos (e.g. G. Zhu et. al. 2006). However, many of those methods are based on heuristic steps such as segmentation and recognition of sport- and scene-specific objects or on trivial low-level image features (e.g. Geetha and Narayanan 2008). For automatic visual recognition, on the other hand, Otsu and Kurita (1988) has proposed a scheme of adaptive vision system which comprises two stages of feature extraction; namely, higher-order local auto-correlation (HLAC) and multivariate analysis. Since Otsu’s proposal, HLAC has been successfully applied as primitive image features in pattern recognition tasks such as object, face, posture recognition and so on. HLAC extracts image features as histograms of local morphological patterns at each frame.

In this study, we apply similar scene detection method using HLAC and multiple regression analysis to various sport videos and demonstrate the usefulness of the present method.

The experiment was performed to examine whether the method was useful for automatic detection of similar scene of interest in sport videos. In this experiment, the HLAC image feature extraction method was applied to wide variety of sport videos under several different conditions, and their performance was evaluated. The difference among the conditions are given by dividing each of input image sequences into more small regions such as 4 regions, 9 and 16 and so on. Each values of HLAC image features can be more sensitive to subtle changes in input images.

For evaluating the performance of the proposed method, multiple-fold cross-validation is used and the resultant precision and recall rates, F values are evaluated.

The averaged detection rate and the recall rate was almost over 95% and the computational time was short enough to allow on-line processing, which exhibits potential ability of the proposed method. It indicates that the HLAC approach can be applicable to efficient data collection on wide variety of sport videos. The approach can make the video processing and handling for sport coaching and PE more efficient and effective and also encourage and facilitate sports practitioner’s interaction.
Keywords:
Higher-order local auto-correlation, similar scene detection, sport coaching.