• No results found

2.4 Computational Gesture Models

2.4.2 Gesture Models in Computer Graphics

2.4.2.1 Expressive Movement Generation

Techniques for the generation of expressive movement can be roughly divided into four categories: (1) adding expressiveness to neutral motions, or providing tools to modify motion expressions, (2) making the existing motions t some constraints, (3) adding secondary movements, and (4) controlling behaviors. This classi cation is made for convenience of the presentation; in practice, these techniques are frequently combined to achieve the best animation results.



Adding expressiveness to neutral motions or providing tools to edit motion

expressions

Several researchers have suggested methods of adding expressiveness to animated motions using such methods as stochastic noise functions [108], Fourier function models [129], signal processing [24], or emotional transforms [2].

Perlin uses rhythmic and stochastic noise functions to de ne time varying parameters that drive animated puppets [108]. The user controls the puppet through a set of buttons, representing a set of primitive actions and discrete states of the puppet. The system can smoothly blend the selected primitive actions into a coherent animation if the relative contribution (weight) of each action is speci ed properly. The user can tune expressions by adding a pseudorandom noise function to joint motions, modifying joint angle frequency and amplitude, and controlling transition times for di erent actions. The noise functions give the e ect of subtle restlessness and weight shifting, adding low frequency \texture" to the motion. The resulting animated puppet is thus in constant motion and appears to have a dynamic, life-like motion quality. However, it is hard to judge the range of expression possible with the system. It seems the scheme only works ne for rhythmic, repetitive actions, such as walking and dancing. Non-rhythmic motions are selected stochastically for variations. Also, varying expressions by modifying the puppet's scalar joint angles over time

t

via sine and cosine functions is non-intuitive and limits movement qualities. Setting transition times and action weight also requires a certain artistry and skill. If these

parameters are applied naively, the resulting animations can be disastrous.

Unuma et al. use Fourier analysis techniques to interpolate and extrapolate human locomotion data to capture a wide variety of expressions [129]. For instance, they can generate various degree of \tiredness" by interpolating between a normal \walk" and a \tired" walk. In addition, by quantifying the di erences between the coecients of a Fourier function model for a neutral locomotion and those for emotion-driven locomotion, they can generate di erent Fourier characteristic functions, which can then be, individually or in combination, applied to other neutral locomotion to produce di erent variations and expressivities. However, the process of generating Fourier functional models and characteristic functions could be very lengthy.

Bruderlin and Williams apply multiresolution ltering techniques from the image and signal processing domain to manipulate the neutral motions by treating motion parameters (such as joint angles and coordinates) as sampled signals [24]. When a motion parameter signal passes through a series of lters, an animator can add an emotional component, exaggerate the movement, or constrain joint ranges by adjusting the amplitudes of high, middle, or low frequency bands appropriately. Witkin and Popovic describe a technique for editing of captured or keyframed motion by warping and blending motion parameter curves [137]. For each motion curve, the animator chooses a few keyframes and modi es their poses using a suitable timewarp function. The modi ed poses serve as constraints on a smooth deformation to be applied to the captured motion. The new motion curve satis es the constraints while preserving the nal details of the original curve. The animator warps each motion curve independently. The motion clips are concatenated using Perlin's blending techniques [108]. A wide range of new realistic motions can be created from a single prototype motion sequence. However, motion warping is a purely geometric technique, not based on any deep understanding of the motion's structure. Some warps may appear unnatural and distorted.

Amaya et al. present a method to derive emotional transforms by taking the di erences between neutral and emotion-in uenced actions [2]. They then apply the derived emotional transforms to neutral actions to generate a wide range of

movements with di erent types of expressivities. In order to express individuals' di erences in gender, age, manner, culture, and personality, this approach may need to store and retrieve a large number of emotional transforms. As the same individual shows di erent emotions under di erent scenarios and internal states, this requires a clever indexing scheme if an emotional transform database is designed. The awkwardness in the manipulation may indicate that emotional transforms, along with the noise functions, Fourier functions and ltering functions, only capture the essence of the movement super cially.



Making the existing motions t some constraints

Witkin and Kass present a spacetime constraint technique to produce the optimal motion which satis es a set of user-speci ed constraints [136]. Cohen develops a spacetime control system which allows a user to interactively guide a numerical optimization process to nd an acceptable solution in a feasible time [33]. Liu et al. use a hierarchical wavelet representation to automatically add motion details [84]. Guenter et al. adopt this approach to generate a smooth transition between motion clips eciently [52]. Gleicher simpli es the spacetime problem by removing the physics-related aspects from the objective function and constraints to achieve an interactive performance [50].



Adding secondary movements

The use of secondary movements has been proposed as a way to enliven animated characters and/or scenes. Although the secondary movements are not the primary focus of the motions of an animated character, their absence can distract or disturb the viewer, making the character unbelievable and unnatural. One approach is to add secondary movements to the primary movements of walking characters based on user- speci ed personality and mood [94]. Another approach focuses on passive motions like the movement of clothing and hair, generated in response to environment forces or the movements of characters and other objects [98]. The secondary movements and the primary movements combined give a richer and more varied set of movements capable of responding to subtle changes in an animated character's personality, manner, and environment.



Controlling behaviors

Expressive movements are also investigated with an aim to build autonomous characters (or creatures) that are endowed with varying behaviors, personalities or goals. The prominent work in this area is that of Reynolds [111], followed by Badler [7], Bates [13], Tu and Terzopoulous [128], Hodgins [55], Thalmann and Thalmann [17], and Hayes-Roth [54]. Although self-animating characters or creatures have demonstrated more-or-less di erent high-level behaviors, their low- level movements are frequently stereotyped, or clumsy and unnatural. In addition, the expressions and their manifestations are usually hard-wired in the code and very in exible to recon gure and extend. Blumberg and Galyean [21] and Funge et al.[48] address these concerns by introducing mechanisms that give the animator greater control to direct autonomous characters to perform speci c tasks, however, their work is at best partially successful, and the impression that one gets from watching even the most recent e ort in making autonomous agents is that their basic movements are still fairly unexpressive, lacking the qualities that make them look \right."

In general, most of these techniques are valuable for generating expressive movement; however, either these methods require an o -line modeling process for each di erent type of expression, or the modi cation process involves nonintuitive low-level manipulations in such a way that some artistry or expertise is demanded in order to generate natural, expressive movements, or both. In addition, they may prove dicult or costly to use in generating the range of expressivity of human communicative gestures.