Movement Classes from Human Motion Data

Kang Hoon Lee, Jong Pil Park, Jehee LeeTrans. Edutainment 2011 · DOI

TL;DR

This paper introduces a novel method to automatically identify distinct movement classes from unlabelled human motion data by unifying segmentation and clustering into a single optimization framework. This approach allows users to define the desired number of movement classes, and the system then finds the optimal way to segment and group the motion data, simplifying the process of organizing and understanding large motion datasets.

What Problem Does This Paper Address?

Working with large amounts of captured human motion data, such as for video games or film, presents a significant challenge: how to efficiently understand and organize this data. Manually sifting through hundreds of dance steps or athletic movements to identify distinct types (like "box step" or "free spin") is incredibly time-consuming and often incomplete. Existing methods typically tackle this by first segmenting the motion into intervals and then clustering these segments. However, this two-phase approach creates a dependency where the quality of segmentation heavily impacts clustering, and tuning parameters for both can be difficult. The paper aims to overcome this by integrating segmentation and clustering into a single, unified optimization process.

Methods

The core of this paper's contribution lies in its unified optimization framework. Instead of separate segmentation and clustering steps, they propose a single process that simultaneously determines the optimal segmentation of motion data and the clustering of those segments into movement classes.

  • Genetic Algorithm: To handle the combinatorial complexity of finding the best segmentation and clustering, the authors employ a genetic algorithm.
  • Genetic Representation: Motion data is represented as binary strings. Each bit in the string corresponds to a potential cut location in the motion sequence. A '1' indicates a cut, and a '0' indicates no cut. This allows for an arbitrary number of segments within a defined range.
  • Fitness Function: The "goodness" of a segmentation (represented by a binary string) is evaluated by a fitness function that balances two key aspects:
  • Compression Ratio (R(C)): Measures how efficiently the motion data can be summarized by the identified movement classes. A lower ratio means shorter segments are needed, indicating better compression.
  • Reconstruction Error (E'(C)): Measures the distortion introduced when replacing each motion segment with the centroid of its assigned class. Lower error means the identified classes accurately represent the original motion. The fitness function combines these two metrics, weighted by a parameter w, to find a balance between conciseness and accuracy.

  • K-medoids Clustering: After a segmentation is proposed by the genetic algorithm, the K-medoids algorithm is used to cluster the resulting motion segments into the specified number of classes. K-medoids is chosen for its ability to handle more general distance measures than K-means.

  • Motion Distance: A pose-to-pose distance measure is used, extended to motion intervals using a simplified uniform time warping approach to avoid the computational cost of dynamic time warping.

Datasets

The paper uses various input motions for experimentation, including:

  • Breakdance motion data
  • Boxing motion data
  • Basketball motion data

The paper does not introduce new datasets but rather demonstrates the effectiveness of their method on existing types of motion capture data.

Key Insights and Ideas

  • Unifying Segmentation and Clustering: The most significant insight is that treating segmentation and clustering as a single, integrated optimization problem, rather than sequential steps, leads to better results and simplifies the user's interaction.
  • Flexibility in Segmentation: Allowing an arbitrary number of segments (within a range) and using a genetic algorithm to explore this vast search space provides greater flexibility than fixed segmentation schemes.
  • User-Centric Control: The ability for the user to simply specify the number of movement classes, without needing to define segmentation parameters, is a major advantage for usability.
  • Visual Summarization: The method provides a concise visual summary of motion data through representative "centroid" poses for each class, making it easier to grasp the content of large datasets.
  • Animation Authoring Interface: The classified movements can serve as building blocks for an intuitive animation authoring interface, allowing users to construct new sequences by arranging these movement classes.

Key Results

  • The proposed method successfully identifies meaningful movement classes from unlabelled motion data.
  • Visualizations (Figure 2) demonstrate the ability to summarize diverse motion types (boxing, basketball) into distinct classes.
  • The system enables an interactive animation authoring interface where users can arrange classified movements.
  • Experimental results show that the fitness score (balancing compression and reconstruction error) typically peaks at an "ideal" number of classes, indicating the effectiveness of the fitness metric. For example, in one experiment, the best score was achieved with 13 classes.

Limitations and Future Work

  • Combinatorial Explosion: The paper acknowledges that the arbitrary number of segments leads to a challenging combinatorial explosion problem, which can limit scalability with very large motion datasets.
  • Hierarchical Extension: As a potential solution to scalability issues, the authors suggest a hierarchical extension where short motion sequences are first processed into many classes, and then these classes are merged into fewer, more comprehensive ones.
  • Animation Authoring System Improvements: Future work includes enhancing the animation authoring system to incorporate path editing, pose editing, and multi-character synchronization.

Summary updated

BibTeX
@article{lee2011movement,
  author = {Kang Hoon Lee and Jong Pil Park and Jehee Lee},
  title = {{Movement Classes from Human Motion Data}},
  journal = {Trans. Edutainment},
  year = {2011},
  doi = {10.1007/978-3-642-22639-7\_13},
}
Figure