260 likes | 390 Vues
This paper presents a model for simulating pedestrian movement and interactions, focusing on reactive path following. It emphasizes collision avoidance among multiple characters and utilizes potential fields to guide movements toward goals while considering obstacles. The approach is informed by previous work on boids, social interaction models, and navigation primitives. A Naïve Bayes classifier is employed to learn pedestrian behavior, optimizing movement decisions through pattern recognition of potential collision scenarios, and enhancing realism in animated character behavior.
E N D
Reactive Pedestrian Path Following from Examples Computer Animation and Social Agents 2003 Ronald A. Metoyer Jessica K. Hodgins
Introduction • Need a system to model the movement of many people walking and interacting • Want to maintain control over the path each individual takes • Hard to deal with collision avoidance with many characters • Easy to use
Previous Work • Reynolds • Boid Model for flocks, schools, and herds • Pedestrian Models • Fluid flow model • Inter-pedestrian interaction models (Helbing and Molnar) • Social interaction based on + and – potential fields • Lane formation in halls, queuing, turn taking
2D Character Intelligence • Exploit fact that humans have to move on a 2D plane (for the most part) • Basic level of intelligence • Reactive path following, obstacles, other pedestrians • Social Forces Model • Reactive control utilizes potential fields • Obstacles are repulsive • Goals are attractive
Modeling • Point mass dynamics • Update equation is: • Where the force fx is obtained from the potential field • dt is the simulation steps • m is the mass of the character • Although goal locations can be specified, it is desirable to allow a definable path for the character to follow • People are experts in drawing a path through a scene in the absence of moving obstacles • Can also be generated through automatic process
Path Diagram • User draws a spline path for character to follow • The path is converted into forces by the following: • Character will attempt to follow the direction of the path, but as it gets more off track, it’ll be pulled back stronger
Direction Primitives • Intelligence model will produce correct 2D animation in terms of obstacle avoidance, but not necessarily natural looking • Alert user to potential collisions and ask how to resolve them • Navigation Primitives • Yield, Cut-in-front, Go-around-right, Go-around-left, No-action • Chosen based on traffic planning research
Direction Primitives (Cont.) • Focus on two tasks a pedestrian performs • Monitoring • Observing other pedestrians in the area to determine their navigational intents • Yielding • Act of adjusting velocity (Magnitude or Direction) to avoid a potential collision
Learning • Use previous direction primitive choices to aid the user in future decisions • Direction Primitive • Feature vector that describes current scene • Is the path around left blocked by other pedestrians or obstacles (Y or N) • Is the path around right blocked by other pedestrians or obstacles (Y or N) • Relative speed of the colliding pedestrian (5) • Approach direction of the colliding pedestrian (8) • Colliding pedestrian’s distance to collision (5) • Pedestrian’s distance to collision (5) • Desired travel direction (3)
Learning (Cont.) • Naïve Bayes Classifier • Five primitives are hypotheses • Seven variables are inputs • Potential collisions are classified into one of the 5 primitives • Advantages • Outperforms neural networks and machine learning algorithms in most real life cases • Disadvantages • Limited by the fact that it can only deal with discrete data
3D Motion Generation • Use motion capture • Create a directed graph of poses to get a probability matrix for transitions from one pose to another
Results • Compared the Naïve Bayes algorithm to actual choices made by users • Claim 72% accuracy as opposed to a random choice which would be 20% naturally • This doesn’t mean much, because all it is really testing is their ability to train a Bayes classifier
Limitations • Requires (utilizes) a lot of human intervention • There is no motion capture data of a person stopped, so it appears the person is spinning around when standing still
Basic System U/I Desired State State Controller Dynamics Torques State State Torques Integrator Renderer State
Cart / Pole • Apply torque to cart’s wheels • Balance pole • Accomplish desired location • Accomplish desired velocity • Extra Credit • Swing-up task
Basic input file for Cart / Pole • language = C • gravity = 0 0 -9.80665 • prefix = cartpole • # cart is a truck-sized object, 20 x 4 x 3 feet = 6x1.5x1 meters • # with car-like density of 170 kg / m^3 • # therefore, truck-like mass of 1800kg = 4000 lbs
body = cart joint = slider jname = pos • mass = 1530 • inertia = 414.37500000 4717.50000000 4876.87500000 • bodyToJoint = 0 0 0 • pin = 1 0 0
# A 300 lb = 136 kg ladder that is roughly # 15 x 1.5 x 0.5 feet = 4.6x.45x.15 meters • body = ladder inboard = cart joint = pin jname = theta • mass = 52.785 • inertia = 0.98971875 93.17652187 93.96829687 • bodyTojoint = -2.3 0 0 • inbToJoint = -3.0 0 0.75 • pin = 0 1 0
Swinger • More complicated simulation of girl on a swing • Hands are rigidly attached to rope • Butt is rigidly attached to seat • You control torques at shoulder, elbow, hips, and knee
Swinger • State machine • Swinging has discrete modes, or states • Define when they begin and end • Define what movements are required for each state
Discrete event simulations • Very important!!! • Each simulation has a simulation timestep, DT • Smaller timestep required for larger forces • Numerical imprecision of integrator • Make sure your simulations are precise by dropping DT by an order of magnitude and confirm behavior is the same