Research On Real-Time Rendering Method For Large-Scale Virtual Crowds

Real-time rendering of large-scale virtual scenes is one of the hottest topics in the field ofcomputer graphics. Up to now, it is possible to realize real-time rendering of large-scale staticmodel (such as terrain), while there're still many challenges for rendering of tens of thousandsof virtual human crowds in real-time. The latter can be used to simulate the animated scene ofvirtual crowds in gigantic stadiums or in the subway for three-dimensional visualized virtualtryout of emergency evacuation in these locations, or used to generate thousands uponthousands virtual spectators or soldiers in order to reduce the cost in film-making or to boostvisual fidelity in various video games. The wide usage in these fields as safety tryout,film-making and games and entertainments gives the real-time rendering method of large-scalevirtual crowds a very good prospect.However, available real-time rendering techniques for crowds have more or less someshortcomings;they either generate low-quality image, or make the control of virtual human ina one-by-one way impossible, or can only support limited number of individuals that can bedisplayed simultaneously in the crowds. With these shortcomings, this paper is dedicated toproviding more general solution to high-quality rendering of large-scale virtual crowds inreal-time. It bases itself on geometrical characteristics of the virtual human, and realizes thegoal by reducing the rendering cost per individual. First, it proposes a level-of-detail algorithmthat can reduce the shape deformation when simplified models move so that the virtual fidelityof the scene is enhanced. Second, it makes improvement on one typical point samplingalgorithm and can generate highly simplified virtual human models. An effective drivingmethod of the result from the improved sampling algorithm is also proposed. The combinationof the latter both can generate tens of thousands of virtual crowds in real-time, which in somedegree overcomes the conflict between driving each virtual human model individually andrealizing very large virtual crowds rendering in real-time. The concrete achievements of thispaper are as follows:(1) The paper proposes a mesh simplification method ad hoc to virtual human models.It is based on the specially designed "ring" constraint, by which it reduces the shapedeformation when the simplified model generated from general algorithms moves, andenhances the visual fidelity for the animated virtual crowds in large-scale scenes.Up to now, the general algorithms have achieved rather good results for simplification ofstatic models;however, because these algorithms take only the static pose as simplificationmetrics and take no account of the moving characteristics of the model, direct application ofthem often creates simplified models which are inclined to shape deformation when they move.This paper simplifies virtual human models with poses in this way: it derives from the Qslimalgorithm;by adding ring constraints in simplification around human joints and other criticalsurfaces, it ensures that model details around these locations are reserved as much as possible,so that the shape deformation from the general algorithm is reduced, and the visual fidelityincreases for large-scale scenes composed of animated virtual crowds.(2) The paper proposes a highly simplified model creation and effective rendering methodbased on the point sampling technique. It makes improvement on both one typicalsampling algorithm and the driving method for the multi-resolution model generatedfrom the algorithm, and realizes real-time rendering of virtual crowds composed of tensof thousands individuals.The traditional algorithm for controlling and rendering of virtual human models can generatelarge-scale virtual crowds using the "hierarchical instantiation" technique, but it requires thatthe crowd be divided into groups, which results in uniform control over each group rather thancontrol over each individual. Besides, the algorithm takes too much into consideration thesmooth transition between animation frames, which can lower the rendering efficiency.The paper adopts a different method to realize the rendering of large-scale virtual crowdswith each individually controlled;on one side, it reduces the geometrical complexity of eachmodel without "hierarchical instantiation" by sampling it from higher level of the samplingdata structure, on the other side, it discards the complex sampling and driving for smoothtransition, and uses the simple and direct per-frame sampling method instead to boost thedriving efficiency. The combination of both can to some degree overcomes the conflictbetween increasing the number of virtual human models in the scene and keeping each oneindividually controlled. Furthermore, other optimization modules such as computationdecomposition and reusing, effective data management et al. are added to the system. It'sproved that the method proposed in this paper can generate tens of thousands of high-qualityvirtual human models in real-time with each individual controlled in the one-by-one way.