Abstract
Practical snowboard instructor training is an iterative process divided into two steps. In the first step, the future snowboard instructors are recorded on video performing on the slope. In the second step, these video recordings are analyzed and discussed with the focus on possible improvements of the future instructor's personal snowboarding style. The future instructors then try to apply the improvements in the next iteration of the first step.
This thesis presents a way of adequately supporting the second step by content-based classification and retrieval of snowboard videoclips.
The theory of snowboarding defines several turn types with different difficul- ties that are practiced step-by-step. Because rhythm and speed are the two main characteristics of different turn types, this thesis explores the feasibility to measure them via motion-detection and investigates how to deal with di- sturbing factors like camera shaking.
The proposed method uses the output of optical flow analysis to compute the duration between two turns and the speed of the turn to classify the types of turns.
The audience in theoretical snowboard lessons is ususally bigger than one person, but everyone needs individual feedback during analysis. As a result it is very important for trainers to be able to quickly present appropriate video samples - either from the same or from another person.
This personalized feedback motivates the second presented method in this thesis. This method employs an established color analysis technique to distinguish which person is shown in the videoclips. The method enables trainers to select individual videoclips for presentation.
In order to evaluate the acquired techniques and developed methods, they are applied on a manually generated test-set of videoclips which were recor- ded during several days of training by this thesis' author. Turn type classification yields good results in computing the average number of frames between two shifts in direction (wide driven carving turns versus fast moving short turns) so 85% percent of videoclips are classified correctly (65% even clearly).
The distinction of videoclips based on depicted persons is highly dependent on scenery and illumination, which disturbs classification results because color-matching fails (classification error-rate rises linearly with the number of analyzed videoclips).
This thesis presents a way of adequately supporting the second step by content-based classification and retrieval of snowboard videoclips.
The theory of snowboarding defines several turn types with different difficul- ties that are practiced step-by-step. Because rhythm and speed are the two main characteristics of different turn types, this thesis explores the feasibility to measure them via motion-detection and investigates how to deal with di- sturbing factors like camera shaking.
The proposed method uses the output of optical flow analysis to compute the duration between two turns and the speed of the turn to classify the types of turns.
The audience in theoretical snowboard lessons is ususally bigger than one person, but everyone needs individual feedback during analysis. As a result it is very important for trainers to be able to quickly present appropriate video samples - either from the same or from another person.
This personalized feedback motivates the second presented method in this thesis. This method employs an established color analysis technique to distinguish which person is shown in the videoclips. The method enables trainers to select individual videoclips for presentation.
In order to evaluate the acquired techniques and developed methods, they are applied on a manually generated test-set of videoclips which were recor- ded during several days of training by this thesis' author. Turn type classification yields good results in computing the average number of frames between two shifts in direction (wide driven carving turns versus fast moving short turns) so 85% percent of videoclips are classified correctly (65% even clearly).
The distinction of videoclips based on depicted persons is highly dependent on scenery and illumination, which disturbs classification results because color-matching fails (classification error-rate rises linearly with the number of analyzed videoclips).
Reference
Lepizh, D. (2010). Visual Information Retrieval : automatisierte Klassifikation von Snowboardclips [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://resolver.obvsg.at/urn:nbn:at:at-ubtuw:1-43256