Modeling of multiple character interactions for motion retrieval and synthesis


Student thesis: Doctoral Thesis

View graph of relations


  • Kai Tai TANG

Related Research Unit(s)


Awarding Institution
Award date15 Feb 2012


3D Motion Capture is the technology that records and digitalizes human motions. It has been used in driving realistic characters in animations and video games, etc. However, it takes a lot of labor and time to capture motions and post-process the artifacts during capture. Especially, people may take more effort to post-process multi-character interactions because the occlusion problem and noise are more severe. Hence, it demands for robust methods to reuse captured motions in database. It is also desirable to generate new interactions by combining single character motions. However, it is challenging to produce a natural interaction by simply putting two characters together because their mutual effects like relative acting directions and timings are needed to be adjusted appropriately, and the styles of the original component motions are preserved. In this thesis, we develop novel methods in emulating logical similarity as well as abstracting the semantic characteristics of captured 3D motions based on the relative distance between joints (i.e. JRD). To provide a sufficient amount of data for motion retrieval, we build a motion dataset by discovering the primitive moves in which long motion clips are segmented and grouped into collections of similar subsequences. In retrieval of single character motions, we propose an Adaptive Feature Selection (AFS) method that is able to select distinguishable features by identifying the characteristics of the motion query. We introduce a Graded Relevance Feedback (GRF) scheme that allows user to provide a more informative feedback with finer relevance levels rather than a hard decision of whether "relevant" or "irrelevant". We model and retrieve multi-character interactions by extending JRD to measure the spatial relationship between interacting characters. Furthermore, we abstract the space-time relationships between interacting characters and introduce a metric that emulates their logical similarities. Finally, we apply this model in synthesizing natural interactions from retrieved single character motions that the local styles of constituting single character motions are preserved. At the last part of the thesis, we introduce various applications implemented with our proposed methods and show their usefulness in real life.

    Research areas

  • Motion, Characters and characteristics, Computer simulation, Computer animation