Research On The Data-Driven Techniques For Crowd Animation

In recent years, crowd animation techniques have received much attention in the Virtual Reality and Computer Graphics community. So far, on the one hand, although we can obtain realistic large-scale crowd animation in the non-real-time applications, for example large-scale crowds consisting of virtual characters have become commonplace in Hollywood cartoon films, it is a long haul work to create it; on the other hand, it is still a difficult problem to synthesize realistic large-scale crowd animation sustaining an interactive frame-rate in the real-time applications. One main reason caused this situation is that the num of affect factors of interaction between individuals became larger as the crowd size increased. It is worth while to note that when the crowd size increased, it is a quality change rather than a quantity change. Another main reason is that power of the consumer-level computer's graphics card is still too weak for rendering large-scale crowd.However, because crowd animation is intensively needed in many fields such as entertainment industry, architecture design, military training, and physical training, a lot of new methods had been proposed to synthesize crowd animation. These methods can be classified into four research directions: 1) virtual environmental modeling for crowd simulation which focused on how to simulate the environment where crowd resided more realistically; 2) crowd behavior simulation which focused on how to simulate the crowd behavior more veritably; 3) crowd motion synthesis which focused on how to synthesize crowd's motion more naturally; 4) real-time crowd rendering which focused on how to real-time render large-scale crowd more quickly.The four research directions almost covered all the important problems in the crowd animation research area. This dissertation had done a series of work in each of the four research directions based on a data-driven approach.In chapter 2, a data-driven macroscopic crowd animation synthesis method using velocity fields was proposed to simulate crowd's macroscopic behavior. This method directed the crowd how to move using sets of representative velocity fields which were extracted from video data of pedestrian movement. This method solved the problem that it usually needed lots of empirical parameters setup and arduous user interaction to synthesize the realistic crowd behavior animation. Experiments showed that this method can synthesize realistic macroscopic crowd animation well with little user interaction, and can be used in many situations.In chapter 3, a hierarchical environmental modeling method was proposed to simulate the environment where crowd reside. Based on this method, the dissertation implemented an environment modeling software which can be used to construct a environment model consisted of geometric level, topological level, path level and cognitive level. The environment model can provide many supports for visual effect, sensing, path planning, decision-making, and interaction et al.In chapter 4, a data-driven method for crowd simulation in urban scene was proposed to simulate the microscopic crowd behavior. This method can be used for adding scalable microscopic crowd animation into virtual urban scene. Firstly, we captured several typical pedestrian movement videos, and extracted state-action data sets from them. These data sets were then integrated into an informed virtual urban model. And finally, a crowd animation synthesis system of pedestrian in virtual urban was established based on the informed virtual urban model.In chapter 5, an interactive complicated human animation synthesis method was proposed to add various complicated human animation into crowd simulation scene. This method was consisted of a motion database constructing method, an improved skinning method, an interactive motion synthesis method, and a human body's adjunct object animation synthesis method. The dissertation taken the case of a computer aided motion design system for Rhythmic Gymnastics to prove this method's validity.In chapter 6, a path-driven crowd motion synthesis method was proposed to synthesize various natural human locomotion motions from crowd's paths which were outputted by previous crowd behavior planner. Firstly, the path to be matched was divided into a series of short curve segments. Then each short curve segment was associated with a mostly matched motion clip stored in a structural Motion State Machine. After that, these motion clips were connected one by one to form the final continuous human motion data satisfying the motion path constrain. The proposed method also allowed user to interactively associate specific motion clip to any curve segment. The proposed method is very suitable for the occasion when the path to be matched is in large scale, or multiple paths are matched simultaneity, or paths intersect frequently.In chapter 7, we give a conclusion of this thesis and discuss the future work.