Unsupervised Motion Pattern Learning for Motion Segmentation

Unsupervised Motion Pattern Learning for Motion Segmentation
Markus Weber, Marcus Liwicki, Gabriele Bleser, Didier Stricker
The 21st International Conference on Pattern Recognition International Conference on Pattern Recognition (ICPR-21), November 11-15, Tsukuba Science City, Japan

Abstract:
This paper proposes a novel method for automated generation of motion segmentation models for full body motion monitoring. The method generates, in an un-supervised manner, a motion template for a dynamic warping approach from a short training sequence, i.e., from very few data. Therefore it first automatically detects motif candidates, i.e. the recurring patterns in the training sequence. Then it uses the detected motifs to construct the model. This novel method is able to automatically find motifs in a multivariate time series and generate a model which is capable of segmenting the series in a real-time system. The technology is evaluated in the context of a personalized virtual rehabilitation trainer application during a clinical study. The novel motion capturing dataset is publicly available.

Unsupervised Motion Pattern Learning for Motion Segmentation

Unsupervised Motion Pattern Learning for Motion Segmentation
Markus Weber, Marcus Liwicki, Gabriele Bleser, Didier Stricker
The 21st International Conference on Pattern Recognition International Conference on Pattern Recognition (ICPR-21), November 11-15, Tsukuba Science City, Japan

Abstract:
This paper proposes a novel method for automated generation of motion segmentation models for full body motion monitoring. The method generates, in an un-supervised manner, a motion template for a dynamic warping approach from a short training sequence, i.e., from very few data. Therefore it first automatically detects motif candidates, i.e. the recurring patterns in the training sequence. Then it uses the detected motifs to construct the model. This novel method is able to automatically find motifs in a multivariate time series and generate a model which is capable of segmenting the series in a real-time system. The technology is evaluated in the context of a personalized virtual rehabilitation trainer application during a clinical study. The novel motion capturing dataset is publicly available.