Getting out of the way: collision avoiding pedestrian models compared to the real world G. Lämmel, M. Plaue Numerical simulation of human crowds is a challenging task, and a number of approaches to simulate pedestrian dynamics on a microscopic level have been established. The aim is to model a realistic, and in particular collision-free, movement of crowds in complex environments. One possible approach to achieve this goal are Cellular Automata (CA). CA models represent the environment with a grid-like structure, where each cell of the grid may contain at most one pedestrian at a time. CA models have often been used for the simulation of evacuation scenarios. A common problem with existing CA models, however, is that they do not model complex wayfinding, since in most CA models wayfinding is implemented via a globally defined potential field. In theory, it would be possible to assign an individual potential field to each pedestrian. This approach, however, would be to complex in terms of computational costs for large scenarios. Other simulation concepts use (discretized) differential equations similar to equations known from the description of Molecular Dynamics (MD). Probably the best known model based on the MD analogy is the social force model. In social force model simulations, each pedestrian has a desired velocity towards a desired destination and adapts her current velocity accordingly. A pedestrian's incentive to avoid obstacles such as other pedestrians is modeled by repelling forces. Force-based models are well understood and have reasonable computational costs. There is a third class of models that try to achieve collision free pedestrian movement in complex environments. These models are based on the so called configuration space obstacle approach and have their foundation in robotics. In this context the configuration space describes all possible locations a pedestrian can reach. Locations that cannot be reached are the so called configuration space obstacles. For this approach the pedestrians and the obstacles in the environment (e.g. walls and other pedestrians) have to be represented as a set of simple polygons. A path through the environment is collision free if the path does not intersect the Minkowski sum of the polygonal obstacles with the polygonal representation of the pedestrian reflected in her reference point. An extension to the configuration space obstacle approach is the velocity obstacle approach. Similar to the configuration space obstacles the velocity obstacles describe all velocities a pedestrian can choose that will lead to a collision at some point in time assuming straight movement and no acceleration of rest of the pedestrians. In the velocity obstacle approach every pedestrian chooses at each point in time a velocity that avoids collision and is close to the desired velocity. This work investigates two approaches for explicit collision avoidance for multi-destination pedestrian crowds simulations in complex dynamic environments. The first approach is an extension to an existing collision avoiding force-based model (Zanlungo et al., 2011). In the force-based model the repelling forces do not only depend on the locations of the obstacles and pedestrians but also on their velocities. When calculating the actual force that affects a given pedestrian the model first calculates the minimal time of closest approach to all other pedestrians and obstacles in the environment. All pedestrians are than projected up to this point in time assuming constant movement. The influencing force on the pedestrian in question is calculated based on this projection and weighted by the inverse of the minimal time of closest approaches times. This leads to a behavior that always tries to avoid the next potential collision under the assumption of constant movement. Like in other force-based models, too, each pedestrian has a desired velocity to which she adapts in a given time if free movement is possible. However, most existing force-based models do not address the important problem of complex wayfinding. In our approach, complex wayfinding is implemented via a navigation graph and a shortest path search: a force that lets a pedestrian move along the navigation graph replaces the simple desired velocity vector. With this model it is possible to simulate pedestrian movements to multiple destinations in a complex environment. The second approach discussed in this paper is based on the velocity obstacle approach (see, e.g., v.d. Berg et all., 2008). But instead of calculating valid velocities during every time step the pedestrians choose valid accelerations preferably close to the desired one. Therefore, we refer to it here as acceleration obstacle model. This approach leads to the same result as choosing a valid velocity directly but lets the model be better integrated with force-based models. This integration is needed since the calculation of the desired velocity and the short-range interaction is still based on forces. Like with our first approach, desired movement is based on a force that lets each agent move along a navigation graph. Both models are tested on data from a real-world experiment conducted by us (see, Plaue et al., 2011). The participants of the experiment have been divided into two groups, and each of the groups was instructed to walk along a given path. These paths were arranged such that the two pedestrian groups intersect at an angle of about 90 degrees. The experiment has been recorded and the individual trajectories have been extracted afterwards from video. In this paper the performance of the force-based model and the acceleration obstacle model are evaluated by comparison with the real-world data in terms of the spatial-temporal distribution of pedestrian densities. To this end, we calculate a local density field via kernel density estimation similar to (Helbing et al., 2007) but with variable kernel bandwidth. A further investigation discussed in this paper is the performance of both approaches in terms of computational costs. At the end of the paper a final appraisal is given which of the models is most appropriate under which conditions. J. van den Berg, M. Lin, D. Manocha: "Reciprocal Velocity Obstacles for Real-Time Multi-Agent Navigation", IEEE International Conference on Robotics and Automation (ICRA 08), New York, N.Y.: IEEE, pp. 1928-1935. D. Helbing, A. Johansson, and H. Z. Al-Abideen: "Dynamics of crowd disasters: An empirical study", Phys. Rev. E 75, pp. 046109 (2007). M. Plaue, M. Chen, G. Bärwolff, and H. Schwandt: "Trajectory extraction and density analysis of intersecting pedestrian flows from video recordings", Proc. PIA11, LNCS 6952, 285-296 (2011). F. Zanlungo, T. Ikeda, and T, Kanda: "Social force model with explicit collision prediction", EPL (Europhysics Letters), 2011, 93, 68005