Expressive Gesture and Body Motion Modeling and Evaluation

Character animation plays a central role in many applications such as embodied conversational agents (ECA), humanoid robots and computer animated films. Key frame animation, as a traditional way of making animation, is laborious and time consuming. This thesis presents several methods for generating expressive gesture and body motions on either ECAs or humanoid robots. The goal of the research is to capture important motion properties and synthesize full body motions of desired qualities requiring minimal user input. To achieve this goal, this thesis develops models for both gesture and body motion.;For gesture, a two layer system, implemented on a Honda humanoid robot, was designed to generate a full range of gesture types from arbitrary text. The system first selects gestures types probabilistically from input text and then aligns matching gesture templates - trajectory curves that define the gesture - with speech. An evaluation of the model's parameters demonstrated that our system could generate motions of different levels of expressiveness, excitement and speech synchronization. As a next step, the "chameleon effect", the tendency of people to adopt the postures, gestures, and mannerisms of their interaction partners [20] was examined, as a way to guide gesture algorithms for human-robot/agent interaction. Evidence was found suggesting that people prefer gestures similar to their own over gestures similar to those of other people.;With respect to body motion, two architectures were proposed to add expressive body motions to gesture animation in order to generate full body animation. The first one is a hybrid system that uses motion capture data and procedural animation to add lower body movement and spine rotation respectively. Naturalness is the most important factor in the design. The second one extends this research to generate stylized body motion. The model takes a statistics-based approach using a small, personalized database of body motion clips. A perceptual study validates that the system is able to capture important aspects of expression. In order to better understand the relationship between gesture and posture, a separate experiment was conducted to evaluate different factors that might affect human's sensitivity to motion realism when motion transplantation is applied. Motion transplantation, transplanting the motion of certain body parts from one clip to another that just contains arm gestures, is a simple yet powerful solution to add variance to body motions. However, it will break the coordination among body parts and potentially generate unrealistic body motions. The experiments found that the presence of posture-gesture mergers, defined as when posture and gesture are integrated into a whole unied movement [59, 60], in the source motion lowers output quality. A distance metric based on motion energy variation was found to be related to people's sensitivity to motion realism.;In summary, this thesis covers several experiments on understanding motion properties and explores methods for gesture and body motion modeling. It represents a step towards creating expressive full body motions for ECAs or Humanoid robots automatically.