Mankyu Sung Scalable, Controllable, Efficient and convincing crowd simulation (2005) Michael Gleicher “I have a bad case of Academic Attention Deficit Disorder” Stephen Chenney Flow Tiles
Overview Related Work Low level (probabilistic action selection) High level (situations and compositions) Results Conclusion Related Future Work Assessment
Main observations: Anonymity in the crowd what instead of who action individual matter only in short time contribution A character is only in a few situations at once
Rules based (Reynolds) Not scalable from authoring perspective Hierarchical (Musse) No complex individual behaviour Physics inspired (Helbing) Limited behaviour and interaction Annotated environment (The Sims, Kallmann)
To select new state evaluate all possible states with behaviour function Default behaviour functions: ImageLookup TargetFind Overlap State: s = {t, p, θ, a, s - ) P k (s) = 1 / (1 + e -αx )
Create complex behaviour by composition of simple behaviours
Situations spatial (ATM, crossing) non-spatial (friendship) When in situation: extend state graph attach sensors add event rules add behaviour functions Composition means union
Tested on 3 scenarios: Street environment crossing street, traffic sign, in-a-hurry Theatre environment horizontal queue, follow, gathering, stay-in... Field environment follow, group, close
1,3 GHz processor 1GB memory 500 agents with increasing number of situations increasing number of agents with 10 situations
Framework can create complex behaviours while minimising data stored in each agent Future work: take into account multi-agent statistics such as crowd density more efficient simulation so not all crowd members go through simulation step at same time explore other mechanisms to combine behaviours to avoid time scale problem
Situation Agents: Agent-based Externalized Steering Logic Schuerman, M., Singh, S., Kapadia, M., Faloutsos P., The Journal of Computer Animation and Virtual Worlds, Special Issue CASA 2010, Wiley, pp. 1-10, 2010, in press. Motion patches: building blocks for virtual environments annotated with motion data Lee, K. H., Choi, M. G., and Lee, J , SIGGRAPH ’06: ACM SIGGRAPH 2006 Papers, 898–906.
Goals clearly specified Situation approach seems to indeed limit the complexity of the agents Problems and possible solutions presented Clearly structured and well written
Claims and assumptions Anonymity justifies probabilistic method? Not for low density crowds People stopping in middle of crosswalk Waiting for traffic light, then not moving when it is green
Implementation details Naive default behaviours Path planning PRM + Dijkstra PRM pre-computed, no dynamic obstacle handling How are states judged to make the character move towards position? Possible local minima? Collision detection No prediction, possible oscillations
Implementation details: extending the state graph extension only with default graph no interaction between situations controlling combination of behaviour functions use of alpha not intuitive, when to use alpha and when to delete a behaviour
Limited experiments maximum of 10 situations maximum of 500 agents random situations added, does this include composite situations?
Impact and applications Limitation on kind of applications no evacuation simulation Situational approach might be a good idea but should be combined with other methods Inspiration for further research