Reactive pedestrian path following from examples

Transcription

1 1 Introduction Reactive pedestrian path following from examples Ronald A. Metoyer 1, Jessica K. Hodgins 2 1 School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, OR 97330, USA metoyer@cs.orst.edu 2 Robotics Institute, Carnegie Mellon University, Pittsburgh, PA , USA jkh@cs.cmu.edu Published online: 18 November 2004 c Springer-Verlag 2004 Architectural and urban planning applications require animations of people to present an accurate and compelling view of a new environment. Ideally, these animations would be easy for a non-programmer to construct, just as buildings and streets can be modeled by an architect or artist using commercial modeling software. In this paper, we explore an approach for generating reactive path following based on the user s examples of the desired behavior. The examples are used to build a model of the desired reactive behavior. The model is combined with reactive control methods to produce natural 2D pedestrian trajectories. The system then automatically generates 3D pedestrian locomotion using a motion-graph approach. We discuss the accuracy of the learned model of pedestrian motion and show that simple direction primitives can be recorded and used to build natural, reactive, path-following behaviors. Key words: Animation Pedestrian simulation Reactive control Three-dimensional (3D) models of architectural and urban designs are increasingly used to visualize expensive construction concepts before full production, to plan complex urban areas, and to visualize pedestrian evacuation patterns. Architects and engineers evaluate options during the design process with 2D simulations and 3D visualizations. Land developers use visualizations to produce compelling marketing graphics and contractors use them to communicate to construction estimators, sub-contractor bidders, and building owners [47]. These visualizations can save time and money for those involved in the construction and urban planning process. To identify potential problems, the scenes should be produced with accurate structure models as well as realistic human inhabitants. Presently, these scenes often do not include human agents because natural human motion is difficult to create. The animation of high-level behaviors of humans is particularly difficult and time consuming to produce. For example, a scene with human characters in a courtyard would require that the user generate locomotion for each character taking into account collision avoidance (Fig. 1). For a scene of even ten characters, this task becomes difficult. Animators and programmers have developed skills and techniques for generating strikingly realistic human characters. Unfortunately, those who wish to generate animated figures are often not experts in animation or computer programming. We are interested in generating natural reactive path planning by building on user expertise in human navigation strategies. In this paper, we develop a system that allows novice users to control the reactive path planning of human characters. We present an approach that computes most of the character motion automatically while still giving the user control over the resulting animation. While the gross character motion is specified by the user via desired paths, the fine details of the navigation, such as obstacle avoidance and path following, are implemented with automatic reactive navigation techniques. The user can refine the motion by directing the characters with navigation primitives. The system uses this direction along with other information about the scene to build a model of desired reactive behavior for use in similar situations. The reactive navigation, usersupplied paths and learned model are combined to produce 2D time-stamped trajectories. We ultimately need 3D motion that tracks the 2D trajectories. We use a motion-graph algorithm that pieces together The Visual Computer (2004) 20: Digital Object Identifier (DOI) /s z

2 636 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 1 2 Fig. 1. Architectural visualization of a crowd scene in a model of the College of Computing at the Georgia Institute of Technology. Models provided by the Imagine Lab at Georgia Tech Fig. 2. In our system, the user s goals are given in the form of desired paths and interactive direction. The interactive direction is used to build a model of desired reactive behavior. This user input is combined with the automatic reactive controller to generate 2D simulation trajectories that are then tracked by 3D characters using a motion-graph approach sequences of frames, or poses, to track the 2D trajectories in time and space while maintaining natural transitions (Fig. 2). We test the learned model quantitatively by performing cross-validation on a set of user direction examples. We also qualitatively compare the resulting 2D motion to that of models using only reactive control and models using random choice. We assess the quality of our 3D motion generation with trajectory tracking experiments. 2 Background Behavior control has been a topic of interest in several fields including computer graphics, robotics, and urban planning. Early interest in the computer graphics community was sparked by the seminal work of Reynolds that introduced the Boids model for flocks, schools, and herds [38]. Recently, Massive Software won an Academy Award for the behavioral animation software used to generate battle sequences with thousands of interacting characters [43]. Several others in the graphics area have made progress in creating realistic behaviors for human-like characters, fish, and dinosaurs [14, 36, 46, 48]. Most of these solutions have focused primarily on autonomous behavior. In 1995, Blumberg introduced a model for directing character behavior at multiple levels, giving the user the ability to control the behavior [8]. More recently, he presented a framework for designing a dog character that learns herding behaviors based on clicker training [7]. Our aim is also to add the human into the behavior control and training loop. In the area of intelligent agents, Barnes takes a more direct approach to character training by designing a virtual environment for interactive visual programming of agents. The user specifies preconditions,

3 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 637 post conditions and the corresponding actions visually within the environment [4]. In other work, we focused on generating natural reactive sports behaviors via demonstration of full trajectories [30]. In this paper, as well as in our previous work, we aim to indirectly collect information about the user s desires in order to build a model of user preferences and produce motions that more closely match these preferences. Several researchers in the graphics area have focused specifically on generating pedestrian behaviors such as navigation planning and natural locomotion [9, 21, 24, 35, 44]. Others have focused on designing simulation environments for generating realistic, directable crowds [12, 32 34, 45]. Ashida and his colleagues analyze actual crowd video to build statistical models of higher level pedestrian behavior [3]. Goldenstein and his colleagues develop an approach grounded in dynamic systems theory to generate reactive behaviors and crowd behaviors for autonomous agents [16]. Pedestrian behavior modeling is also of interest in urban planning. Urban planning researchers build models of human behavior that are valuable tools for planning and design of urban areas such as shopping malls and city centers [5, 6, 10, 15, 27]. Pedestrian models have been developed at several levels of granularity ranging from coarse fluid flow motion models to fine grain inter-pedestrian interaction models such as that proposed by Helbing and Molnar [17 19]. Helbing s model of social interactions is based on attractive and repulsive potential fields very similar to those used by the roboticists for mobile robot control [2, 20, 23, 28]. Using their social forces model, Helbing and Molnar are able to recreate phenomena such as lane formation in halls, queuing, and turn-taking at doorways. More recently, Quinn and Metoyer implemented a parallel version of the social forces model that can simulate up to 10,000 pedestrians in real-time [37]. Others have used cellular automata approaches for modeling pedestrian motion at a similar granularity and have observed similar emergent behaviors [6, 10, 39]. Each of these approaches generates motion that is particularly useful for observing patterns and collecting statistics, but they typically generate roboticlooking motion that is not suitable for realistic visualization. Boston Dynamic s PeopleShop system also provides an interface for novices to design pedestrian content for 3D training and visualization scenes. They allow users to specify paths, sensor regions and behavior for characters in a terrain and allow for run-time control of characters using a joystick. Rather than design fully autonomous characters they implement intelligence amplification to build on the user s intelligence [11]. We are also interested in simplifying the content creation process for novices by providing simple interfaces and characters that improve performance with time. 3 2D Intelligence model To populate a 2D scene with animated pedestrians, the user first describes the motion on a 2D floor plan of the scene to be animated. To relieve the user of some of the tedious details involved with human navigation, we provide a low-level character intelligence model. Character intelligence provides a basic level of behavior for the character via reactive path following in the presence of obstacles, desired paths, and other pedestrians. Reactive control using potential fields is a well studied area in mobile robotics and in other fields such as pedestrian modeling. We use the social forces model of Helbing and Molnar to model the reactive intelligence of a pedestrian in a 2D representation of an architectural environment [18]. This approach defines obstacles as repulsive potentials, goals as attractive potentials, and combines all potentials to produce a composite potential field (Fig. 3). In 2D space, each pedestrian is modeled with point mass dynamics. The update equation for a point mass is x t+1 = x t + ẋ t dt f x m dt2, (1) where the force f x is obtained from the behavior control potential fields, m is the character s mass, and dt is the simulation time step. Similar equations hold for y. The velocity of each point-mass pedestrian is clamped at a limit of 2.0m/s to allow for fast walking. The social forces model described above provides the user with the ability to specify desired goal locations. In order to allow for more control over the path to the goal location, we provide the user with the ability to supply a natural path. People, as experts in navigating physical environments, can visualize and draw natural paths between two points in a scene

4 638 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 3a 3b 3c 3d 4 Fig. 3. a and b Attractive and repulsive fields centered on the circle. c A repulsive pedestrian field moving in the positive x direction. d A boundary field caused by the wall Fig. 4. Snapshot of the user interface. The dashed lines represent boundaries that the pedestrians are aware of. The squares represent obstacles. The user supplied natural path is the general path the user wants the pedestrian to take, while avoiding all collisions along the way. The path is only shown for the currently selected pedestrian in the absence of moving obstacles. The user supplies these paths by drawing directed lines across the floor plan of the environment (Fig. 4). These paths can also be generated using global path planning algorithms such as visibility roadmaps [25]. The paths provide general path guidelines while the reactive control accounts for potential collisions along these paths. The user-supplied paths are converted into forces: f path = k p ( p d max f (d max) ( + 1 p ) d max ) f (s), (2) where f (s) is a unit vector along the path starting at the nearest point on the path, f (d max) is a unit vector perpendicular to f (s), p is the perpendicular

5 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 639 Fig. 5. Force diagram for a character at several points in time along the user-specified natural path. At t A the character is on the path. At t B the character has been forced off course and must return to the path. As distance from the path increases (at t B ), the force direction begins to point towards the path. When the distance is small (at t A ), the force direction is aligned with the path distance from the nearest point on the path, d max is the largest allowed perpendicular distance, and k p is a path gain. The force on the pedestrian is based on the pedestrian s perpendicular distance from the spline that represents the desired natural path that it is following (Fig. 5). These reactive models provide a simple system for the user to populate an environment with pedestrians and direct them from one point to another within the environment. 4 Behavior models from direction primitives The intelligence model alone shouldproducecorrect 2D motion in terms of avoiding the defined obstacles and reaching goals, but will not necessarily produce natural motion for navigating complex scenes. For example, the characters will typically proceed along a path until a repulsive force causes a sudden direction change, while in real life, path adjustments to avoid collisions are often very subtle (Fig. 6). When the motion resulting from the intelligence model does not meet the user s goals, the user is able to interactively direct the character with navigation primitives. As a 2D pedestrian simulation progresses in time, potential collisions may arise and the user is visually alerted to the situation. The user can then stop the simulation and provide direction from the following set of navigation primitives: yield, cut-in-front, go-around-right, go-around-left, and no-action. The user observes the motion and accepts the direction and continues in the animation process or the user revises the direction. The set of navigation primitives was chosen based on research from the traffic planning field. There are five main tasks a pedestrian undertakes while navigating: monitoring, yielding, checkerboarding, streaming, and avoiding perceptual objects [13]. Checkerboarding occurs in a lane when one person following another person does so at a slightly offset position to the left or right in order to avoid stepping on the person in front (and to see around), creating a checkerboard or zipper pattern of pedestrians. Streaming is the act of following directly behind someone in a crowded situation as they create a clear path. Avoiding perceptual objects refers to the avoidance of something that is not truly a physical object that could cause a collision. For example, lines on the pavement are considered perceptual objects. In this paper, we are concerned with two of these tasks, monitoring and yielding, because these are relevant to pedestrian-pedestrian collision avoidance. Monitoring refers to the act of observing pedestrians in the nearby area to determine their navigation intentions. Yielding refers to the act of adjusting velocity (magnitude or direction) in order to avoid a potential collision. The yield primitive is designed to alter the velocity of the pedestrian slightly so that it allows another pedestrian to pass safely in front. The system chooses a target point along the pedestrian s desired natural path before the predicted collision point. A new desired velocity is computed that will put the pedestrian at this point at the predicted time of impact (Fig. 7). Once the collision danger has passed, the pedestrian resumes his original desired velocity. The cut-in-front primitive is implemented in a similar manner, choosing a target point ahead of the predicted point of impact. The go-around primitive is designed to generate an alteration to the desired path that takes the pedes-

6 640 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples Fig. 6. The social forces model alone produces interactions that are often robotic and unnatural. In this example, the characters get very close before taking actions to avoid the collision. Although this type of near miss does happen in real life, it is much less common than what we see with the social forces model, especially in uncrowded situations Fig. 7. The circles with arrows represent pedestrians and their velocities. The yield navigation primitive chooses a velocity that will allow the pedestrian to have a 0.75 m cushion before the potential collision. The cut-in-front primitive chooses a velocity that will give the pedestrian a 2.25 m cushion beyond the potential collision spot Fig. 8. The circles with arrows represent pedestrians and their velocity. The colliding pedestrian is shown approaching from both the left side and the right side (A and B) to demonstrate the two computed avoidance targets for the around-right primitive. The around-right navigation primitive chooses a desired path offset based on the approach angle of the potentially colliding pedestrian. If the pedestrian is approaching from the left, as in A, the offset is chosen with respect to the collision point on the desired path. If the pedestrian is approaching from the right, as in B, the offset is chosen with respect to that pedestrian himself trian around the potential collision spot and back to the natural path. The path around the collision depends on the relative travel direction of the other pedestrian. An offset position is chosen based on the other pedestrian s approach angle and a linear path through this position and back to the natural path makes up the path adjustment (Fig. 8). A similar procedure produces a path for going around to the left. 4.1 Generalizing direction examples One of our goals in this work is to ease the burden on a novice animator who is designing a scene. Rather than discard the novice user s direction examples, we use them to guide the character in future situations. Generalization of direction primitives requires that the system not only record the direction primitives themselves, but also the features of the scene. The

7 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 641 feature vector describes the aspects of the scene that affect the path planning of the character and computation of these features represents a pedestrian s monitoring behavior. Ideally, features are fully descriptive of the situation while easy and fast to compute. A pedestrian s motion at any point in time may depend on several obstacles or several other pedestrians as well as the architectural situation (wide or narrow hall, etc.). We have chosen a set of seven discrete features to describe the situation of a pedestrian in an urban scene. The seven features are: 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). The numbers in parentheses represent the number of discretized values for each feature. Speed and direction are clearly important when trying to negotiate a collision. The blocked right and blocked left features allow us to account for potential collisions aside from the one under consideration. A single value for each of these variables makes up a feature vector. We experimented with both a naive Bayes approach and decision tree approach for learning the behavior model from user examples. The four primitives described above (around-left, around-right, yield, letpass) and a fifth, no-action primitive, are used as the choices or hypotheses for the naive Bayes classifier while the seven variables above make up the features used as input to the naive Bayes classifier. We treat potential collision situations as a choice. The system must classify potential collisions as one of five possible alternatives represented by the direction primitives. The naive Bayes classifier is defined as H max = argmaxp(h i ) h i H j P(a j h i ), (3) where h i are members of H, the set of five possible actions and a j represents the set of seven attributes, or features, described above. The probability of a particular hypothesis, h i is P(h i ) = n h i N, (4) where N represents the number of examples seen thus far and n hi represents the number of examples where theith primitive was chosen. The conditional probabilities for the attributes can be computed by counting the occurrences of each attribute given each particular hypothesis. To avoid a biased underestimate, we compute the m-estimate for estimating the conditional probabilities [31]. We have also experimented with the C4.5 decision tree algorithm for modeling the decision making process of the pedestrian characters [31]. Decision trees are particularly appealing because we have discrete valued primitive choices. Decision trees are also known to be robust to errors in the training data. This property is desirable because the user may give a direction example that contradicts a previous direction example. Decision trees are also interesting because they can be represented as sets of if-then rules resulting in human-readable rules for pedestrian reactive behavior. In the decision tree, the nodes represent tests of one of the seven attributes or features described above. A node has a child for each possible value of the feature. The leaf nodes represent decisions. A decision tree classifies an instance by sorting the instance down the tree to a leaf node. Each path to a leaf represents a conjunction of feature tests and the entire tree represents a disjunction of these conjunctions. The tree is built by a top-down greedy search that chooses the best feature to test at the root and each subsequent level. Best is defined as that which results in the most information gain [31]. We will discuss the results of both the naive Bayes and the decision tree model in Sect Run-time behavior Using the naive Bayes probabilistic model we can now compute the probabilities of the five possible decisions or hypotheses: go-around-left, go-aroundright, yield, cut-in-front, and no-action. The hypothesis with the highest probability is chosen. When using the decision tree approach, we can simply provide the features to the algorithm, which then produces a single output representing one of the five

8 642 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples decision choices. The model can be used in later situations to reduce the amount of direction necessary from the user. At run time, the pedestrians compute potential collisions by observing the velocity of nearby pedestrians to determine future locations. When a potential collision arises, the pedestrian consults the behavior model to determine the proper navigation action to take. The chosen action runs to completion or until another possible collision occurs. The reactive intelligence model is still used to accountfor any collisions not avoided by the learned model and to track the desired path. Both the learned model and the reactive model produce forces that are summed together to generate a final force for the pedestrian. 5 3D Motion generation For visual display, we need 3D motion that tracks the trajectories from the 2D simulation. We first capture multiple locomotion sequences such as walking straight at a comfortable and slow speed, turning by various amounts, starting and coming to a complete stop. Due to limitations of our motion capture system, we were able to capture only short segments of motion in a 9 9 foot capture region. We use a motion-graph-based approach to order the motion capture sequences [1, 22, 26, 29, 40 42]. First, we build a transition matrix that defines a distance, in pose space, between any two poses in our set of motion sequences. We then use a beam-search combined with a cost function to determine the sequence of poses that tracks the trajectory while maintaining natural transitions. 5.1 Motion capture data We use approximately 22 motion capture sequences to generate the 3D trajectory tracking motion for each pedestrian in the urban scenes. The sequences are sampled at 30 Hz and stored as individual frames of motion or poses. A pose is the set of 14 joint angles and the global position and orientation of the root node, the pelvis (Fig. 9). The yaw information of the root node is extracted so that global orientation is defined without respect to facing direction on the ground plane. The poses are stored along with relative position and orientation offsets of the root body between that pose and the previous pose. The offset information allows us to begin any sequence at an arbitrary location and Fig. 9. A single frame of motion capture data is stored as a pose. Our pose consists of 14 joint angles as well as the position and orientation of the root body and pelvis orientation on the ground plane. The offsetpositions are computed as p = p i p i 1,wherep i is the root body s world translation for pose i. The offset orientation is computed as the included angle between two successive global orientations of the root body, M = M T i 1 M i. Both the offset angle of the root body and joint angles can be represented as a vector with a direction that represents the axis of rotation and magnitude that represents the amount of rotation. A single pose is described by 16 vectors: 14 joint angles, the root body orientation offset, and the root body translation offset. A single motion capture sequence is replayed by placing the character at a position and yaw angle on the ground plane and subsequently placing consecutive poses of the sequence. The root position and orientation are updated with the offsets. The joint angles are used directly for each pose. We can now reorder the poses to produce new sequences of walking motion. The goal is to produce a sequence of poses that takes the pedestrian along a predefined 2D trajectory while maintaining natural transitions from pose to pose. 5.2 Pose transition matrix To determine a resequenceing of the motion capture pose data that will track the 2D trajectory while making smooth transitions, we need a measure of the quality of a transition from one pose to another. We first compute the distance between poses i and j

9 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 643 N D ij = p ik p jk, (5) k=0 where N is the number of joints in the pose and p i is a vector representing the ith joint angle. Next, we compute a transition matrix that holds the cost of transitioning from any pose to any other pose of the motion capture data: T ij = 0.5D i, j D i+1, j. (6) This transition matrix holds the transition costs for every pose of motion capture data to every other pose, regardless of which sequence (straight, slow, curved, etc.) the pose came from. Transitions with values above a threshold are pruned. After pruning, the matrix may contain dead ends, or poses from which there are no reasonable transitions. Transitions to dead ends are given high values. Figure 10 shows the transition from a straight walking sequence to a turning sequence. 5.3 Transition searching Once the transition matrix has been computed, generating the trajectory tracking motion becomes a search problem on the graph represented by the transition matrix. The problem is to find the path through this graph that produces natural transitions (small transition costs) while also producing a trajectory for the root body that minimizes distance from the desired 2D trajectory in space and time. To include this 2D trajectory tracking error in the transition computation, we define a cost function Cost ij = w t T ij + w p Error tra j, (7) where w t and w p represent weights for the transition and trajectory error components and Error tra j represents the trajectory tracking error of the root body (pelvis) when taking the transition from i to j. We use a beam search that is similar to a breadth first search except that at each step, it keeps only the m best paths through the graph to this point and discards the others. This algorithm will not produce optimal paths but is a tractable approach that produces reasonable paths. Figure 11 shows the result of a scene in which one character chooses to avoid the collision by going around the other pedestrian to the right Fig. 10. The top character is walking along a straight path. The bottom character is walking a curved path. Imagine a trajectory in which the character must walk straight for 10 m, turn right, and walk straight again. The character must choose a point at which to transition from the straight sequence to the turning sequence in order to minimize error in tracking as well as maintain smooth motion transitions Fig. 11. This scene shows two characters initially on a collision course. In the 2D simulation, one pedestrian chose to go-around-right in order to avoid the collision. This decision resulted in a 2D trajectory different from its desired path shown with the arrow. The path adjustment is subtle but effective in avoiding the collision. The resulting 3D motion consists of poses from several example navigation sequences including walking straight, turning right, and turning left in order to get back to the original desired path. (To view the movie for this scene, see 6 Results We measure performance of the system both quantitatively and qualitatively. A user provided a total of 146 direction examples over several scenes. Each example consists of the feature vector for that particular example and the user s specified action. The

10 644 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples Fig. 12. Filmstrip of the tracking performance of four 3D pedestrian characters. Images are viewed from left to right and top to bottom. The spheres represent the character s desired location every 10 frames. The motion of each character spans approximately 10 meters. The resulting motion is a sequence of poses from several motion capture sequences where each sequence spanned no more than approximately 4 m. Transitions between poses are determined automatically by the search over the transition graph learned model from each scene was carried over to the next scene. To evaluate the generalization of the learned behavior model, we compute a 10-fold cross validation over the entire data set of featureaction pairs. A correct classification is one which produces an action (go-around-left, yield, etc.) that agrees with the user s desired input for the particular example feature vector being classified. We perform this test for two learning algorithms: a naive Bayes classifier and the C4.5 decision tree algorithm. A random guess of the correct navigation primitive would result in 20% accuracy. In our experiments, the naive Bayes classifier produced a 76% accuracy rate while the C4.5 decision tree algorithm outperformed naive Bayes, resulting in 92% accuracy rate. We believe that the success of the decision tree is due to the robustness to contradictory input from the user. It is clear from observation of our data that users do provide conflicting examples, especially in navigation situations where two possible primitives will result in a reasonable reaction. For example, in some situations, it may be perfectly natural to either yield or go-around-left, while in other situations, one primitive is clearly a better choice than the other. We ran several tests to determine the accuracy of the 3D tracking and the transitions (Fig. 12). As curvature increases, tracking accuracy decreases because we have fewer good examples of curved walking than we have of straight walking. Table 1 shows the average pose error, path error and combined error for various trajectories with curvatures ranging from straight to a corner turn (turn 3). Our tests were run with approximately 3500 frames, or two minutes of Table 1. Errors for the 3D trajectory tracking pedestrian. The path and pose errors are computed for various trajectory curvature values ranging from straight to a corner turn (turn3). Pose error represents the average transition cost (Eq. 6) for the duration of the motion. The path error represents the average distance, in meters, of the character s root node (pelvis) from the time-stamped samples of the path for the duration of the motion. In general, as curvature increases, both pose and path errors increase Trajectory tracking errors Curvature straight turn1 turn2 turn3 Pose Path Total

11 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 645 We also qualitatively compared the motion of a crowd using the learned model to a crowd using a random choice model and to a crowd using just a reactive model. In the learned model scene, all pedestrians used the probabilistic naive Bayes model combined with the reactive model to navigate the environment. In the random scene, all pedestrians made random primitive choices to navigate the environment. In the reactive control scene, all pedestrians used only the social forces model to navigate the environment. The pedestrians using the learned model produced much more natural motion than the random choice pedestrians and more natural motion than pedestrians using only the social forces navigation model (Fig. 13 and Fig. 14). 7 Discussion Fig. 13. In these images, the pedestrian s velocity vector has been extended for clarity. A pair of white vectors indicate that two pedestrians are in danger of colliding while a pair of gray vectors indicates that two pedestrians are on a less-urgent collision course. Black vectors indicate no apparent danger of a collision. The top image shows a snapshot of a simulation using only the social forces model to guide the motion of the pedestrians. The middle image shows the same point in time for a system using a random choice of navigation primitives for potential collisions. The bottom image represents a simulation using the social forces model combined with the learned naive Bayes probabilistic model. (To view the movie for this scene, see motion data. The lack of data results in errors in tracking as well as poor transitions in some cases. We believe that with more curved walking sequences, the algorithm performance would improve tracking and transitions for curved trajectories. We have used our approach to animate pedestrians in a 3D visualization of an architectural scene with simple user input while maintaining user control (Fig. 1). In 2D, we allow the user to populate and direct a simple representation of pedestrians in an architectural scene. We further utilize the user s direction to build a model of character behavior for future similar situations. In 3D, we have generated automatic trajectory tracking for articulated 3D pedestrian characters. Although we have chosen a small set of direction primitives, we are able to demonstrate the utility of using this direction rather than discarding it after the animation is produced. When the user must animate a new scene with similar conditions, the model will, in many cases, produce the correct behavior, reducing the number of times the user will have to provide direction. We have demonstrated the utility of learning the desired behaviors using both the naive Bayes and C4.5 decision tree learning algorithms. Our examples have typically shown pedestrian characters in a scene, allowing us to show a reasonable sized crowd that is not too congested. Without congestion, the pedestrians enter situations where they must make strategic decisions about navigating. As scenes get more congested, fewer options exist and navigation strategy becomes more reactive. In our experience, the social forces model alone produces fairly natural results for large scenes with more congestion because most motion is reactive within a very small planning area.

12 646 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples Fig. 13. A filmstrip of three simulations at the same points in time. In these images, the pedestrian s velocity vector has been extended for clarity. A pair of white vectors indicate that two pedestrians are in danger of colliding while a pair of gray vectors indicates that two pedestrians are on a less-urgent collision course. Black vectors indicate no apparent danger of a collision. The images on the left (from top to bottom) are snapshots of a simulation as it runs using only the social forces model for control. The middle images are snapshots of the simulation when making random choices for the navigation primitives. On the right are snapshots of the simulation using the learned behavior model. (To view the movie for this scene, see

13 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 647 A useful result of our 3D motion approach is the automatic choice of proper sequences for velocity changes as well as path curvature changes. For example, if the 2D simulated character comes to a stop in order to avoid a collision and then continues along its path, the 3D algorithm produces a sequence of motion capture poses that brings the 3D character to a complete stop and eventually back to a forward walking sequence. The motion data and search structure enforce natural transitions because the actor who was recorded performed only natural motions and transitions between motions are only made between similar poses. Our 3D motion could be further improved by collecting more data. In the 3D crowd scene, for example, the animator created a situation where two characters met in the middle and appeared to be conversing. Unfortunately, we had no motion capture data of a character standing still, rather, the closest motion to standing still was a sequence of a person turning in place. Therefore, the characters appear to turn their backs on one another. This problem could be solved by adding relevant sequences to the database of motion. More data would also allow us to vary the motion between pedestrians. Currently, all pedestrians draw from the same motion library. We are currently investigating several related research projects. First, we have implemented a parallel version of Helbing s social forces model that can simulate up to 10,000 pedestrians in real-time on a multicomputer. We are also building a tabletop sketch-based interface for interacting with this simulation in real-time as it executes. We hope to use this table-based interface to provide a tool for domain experts, such as a security specialist, to explore what-if scenarios for important situations such as evacuation during an emergency. Our goal is to develop tools and techniques that allow novice users to produce compelling scenes of pedestrians in 3D environments for use in interactive training, visualization, and entertainment. Acknowledgements. The authors would like to thank Dr. Peter Molnar and the Georgia Tech College of Computing for the valuable input and support during the course of this work. References 1. Arikan O, Forsyth DA (2002) Interactive motion generation from examples. ACM Trans Graph 21: Arkin R (1987) Motor schema based navigation for a mobile robot. In: Proceedings of the 1987 IEEE Conference on Robotics and Automation, pp Ashida K, Lee S, Allbeck J, Sun H, Badler N, Metaxas D (2001) Pedestrians: creating agent behaviors through statistical analysis of observation data. In: Proceedings of Computer Animation, pp Barnes C (2000) Visual programming for virtual environments. In: Proceedings from 2000 AAAI Spring Symposium on Smart Graphics 5. Batty M, Jiang B, Thurstain-Goodwin M (1998) Local movement: Agent-based models of pedestrian flow. Center for Advanced Spatial Analysis Working Paper Series (4) 6. Blue VJ, Adler JL (2000) Cellular automata model of emergent collective bi-directional pedestrian dynamics. In: Artificial Life VII: Proceedings of the Seventh International Conference on Artificial Life, pp Blumberg BM, Downie M, Ivanov Y, Berlin M, Johnson MP, Tomlinson B (2002) Integrated learning for interactive synthetic characters. ACM Trans Graph 21(3): Blumberg BM, Galyean TA (1995) Multi-level direction of autonomous creatures for real-time virtual environments. In: Proceedings of SIGGRAPH 95, Annual Conference Series. Addison-Wesley, Boston, Choi MG, Lee J, Shin SY (2003) Planning biped locomotion using motion capture data and probabilistic roadmaps. ACM Trans Graph 22(2): Dijkstra J, Timmermans H (2000) Towards a multi-agent system for visualizing simulated behavior within the built environment. In: Proceedings of Design and Decision Support Systems in Architecture and Urban Planning Conference: DDSS 2000, pp Boston Dynamics (2000) Peopleshop Farenc N, Musse SR, Schweiss E, Kallmann M, Aune O, Boulic R, Thalmann D (2000) A paradigm for controlling virtual humans in urban environment simulations. Appl Artif Intell 14(1): Feurtey F (2000) Simulation of collision avoidance behavior for pedestrians. Dissertation, The University of Tokyo, School of Engineering 14. Funge J, Tu X, Terzopoulos D (1999) Cognitive modeling: knowledge, reasoning and planning for intelligent characters. In: Proceedings of SIGGRAPH 99, Annual Conference Series. Addison-Wesley, Boston, pp Gipps GP, Marksjo B (1985) A micro-simulation model for pedestrian flows. Math Comput Simul 27: Goldenstein S, Large E, Metaxas D (1999) Non-linear dynamical system approach to behavior modeling. Vis Comput 15(7/8): Helbing D (1992) A fluid dynamic model for the movement of pedestrians. Complex Syst 6: Helbing D, Molnar P (1995) Social force model for pedestrian dynamics. Phys Rev 51(5): Henderson LF (1974) On the fluid mechanics of human crowd motion. Transp Res 8: Khatib O (1986) Real-time obstacle avoidance for manipulators and mobile robots. Int J Robot Res 5(1): Ko H, Cremer J (1996) VRLOCO: real-time human locomotion from positional input streams. Presence 5(4):

14 648 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 22. Kovar L, Gleicher M, Pighin F (2002) Motion graphs. ACM Trans Graph 21: Krogh B, Thorpe C (1986) Integrated path planning and dynamic steering control for autonomous vehicles. In: Proceedings of the IEEE Conference on Robotics and Automation, pp Kuffner JJ (1998) Goal-directed navigation for animated characters using real-time path planning and control. In: CAPTECH 98: Workshop on Modelling and Motion Capture Techniques for Virtual Environments. Springer, Berlin Heidelberg New York, pp Latombe J (1991) Robot motion planning. Kluwer, Norwell, MA 26. Lee J, Chai J, Reitsma PSA, Hodgins JK, Pollard NS (2002) Interactive control of avatars animated with human motion data. ACM Trans Graph 21: Lovas GG (1993) Modeling and simulation of pedestrian traffic flow. In: Modeling and Simulation: Proceedings of 1993 European Simulation Multiconference 28. Lyons D (1986) Tagged potential fields: An approach to specification of complex manipulator configurations. In: Proceedings of the IEEE Conference on Robotics and Automation, pp Metoyer R (2002) Building behaviors with examples. Dissertation, Georgia Institute of Technology 30. Metoyer R, Hodgins JK (2000) Animating athletic motion planning by example. In: Proceedings of Graphics Interface, pp Mitchell T (1997) Machine learning. McGraw-Hill, New York 32. Musse SR, Garat F, Thalmann D (1999) Guiding and interacting with virtual crowds in real-time. In: Proceedings of Eurographics: Computer Animation and Simulation 99. Springer, Berlin Heidelberg New York, pp Musse SR, Thalmann D (1997) A model of human crowd behavior: Group inter-relationship and collision detection analysis. In: Proceedings of Eurographics: CAS 97 Workshop on Computer Animation and Simulation, pp Musse SR, Thalmann D (2001) Hierarchical model for real time simulation of virtual human crowds. IEEE Trans Vis Comput Graph 7(2): Noser H, Renault O, Thalmann D, Magnenat-Thalmann N (1995) Navigation for digital actors based on synthetic vision, memory and learning. Comput Graph 19(1): Perlin K, Goldberg A (1996) Improv: a system for scripting interactive actors in virtual worlds. In: Proceedings of SIG- GRAPH 96, Annual Conference Series. Addison-Wesley, Boston, pp Quinn M, Metoyer R (2003) A parallel implementation of the social forces model. In: Proceedings of the Second International Conference in Pedestrian and Evacuation Dynamics, pp Reynolds CW (1987) Flocks, herds, and schools: a distributed behavioral model. In: Proceedings of SIGGRAPH 87, Annual Conference Series. ACM Press, New York, NY, USA, pp Schadschneider A (2002) Cellular automaton approach to pedestrian dynamics. In: Pedestrian and Evacuation Dynamics, Conference Proceedings, pp Schödl A, Essa I (2000) Machine learning for video-based rendering. In: Advances in neural information processing systems, vol 13. MIT Press, Boston 41. Schödl A, Essa I (2002) Controlled animation of video sprites. In: Proceedings of the First ACM Symposium on Computer Animation, pp Schödl A, Szeliski R, Salesin D, Essa I (2000) Video textures. In: Proceedings of SIGGRAPH 00, Annual Conference Series. Addison-Wesley, Boston, pp Massive Software (2004) Massive Sun H, Metaxas DN (2001) Automating gait generation. In: Proceedings of SIGGRAPH 01, Annual Conference Series. ACM Press, New York, NY, USA, pp Tecchia F, Loscos C, Conroy R, Chrysanthou Y (2001) Agent behavior simulator(abs): A platform for urban behaviour development. In: Proceedings of Games Technology Conference Tu X, Terzopoulos D (1994) Artificial fishes: physics, locomotion, perception, behavior. In: Proceedings of SIG- GRAPH 94, Annual Conference Series. ACM Press, New York, NY, USA, pp VISARC (1999) Professional visualization services for the building industry Webber B, Badler N (1995) Animation through reactions, transition nets and plans. In: Proceedings of the International Workshop on Human Interface Technology Photographs of the authors and their biographies are given on the next page.

15 R.A. Metoyer, J.K. Hodgins: Reactive pedestrian path following from examples 649 RONALD METOYER is an assistant professor in the School of Electrical Engineering and Computer Science at Oregon State University. In 2001 he received an NSF CAREER Award for research in computer animation. He currently leads the Interactive Graphics and Vision Lab (IGVL) along with his colleague, Dr. Eric Mortensen. His research focuses on creating interactive spaces for training and education. In particular, he investigates techniques for creating believable character motion and for making animated characters accessible to the novice user. Current projects include motion-capture-based locomotion, content-creation interfaces for novice users, and behavioral control. JESSICA HODGINS is an associate professor of Computer Science and Robotics at Carnegie Mellon University. From she was on the faculty of the College of Computing and the Graphics, Visualization and Usability Center at Georgia Tech. She received an NSF Young Investigator Award, a Packard Fellowship and a Sloan Foundation Fellowship. She was editor-in-chief of ACM Transactions on Graphics from and papers chair of SIGGRAPH Her research focuses on the coordination and control of dynamic physical systems, both natural and human-made, and explores techniques that allow robots and simulated humans to control their actions in complex and unpredictable environments. Ongoing projects include data-driven animation, simulation of human motion, animation interfaces for naive users, and measurements of human perception of animated motion.