A Unified Framework for Evaluating Microscopic Pedestrian Simulation Models

Microscopic simulation models are used in many applications for predicting pedestrian flows with high granularity. They have proven to be a valuable tool to support the design and evaluation of architectural plans, to estimate necessary capacities, to increase safety, efficiency and comfort in crowded areas, and to analyze scenarios for emergency evacuations. Although in the scientific literature a multitude of microscopic models is available, currently implemented simulators typically do not allow for easy and quick switching between models nor do they give insight in their functionality and implementation details. Moreover, there is still a lack of reliable human movement data, which however is a prerequisite for developing models that are able to represent realistic pedestrian behavior as well as for performing model calibration and validation. These shortcomings inhibit to evaluate the capabilities of different models. As a consequence, quantitative comparisons between various approaches are still rare. This doctoral thesis develops for the first time a unified framework for the structured investigation on strengths and weaknesses of different microscopic pedestrian movement simulation models based on an empirical benchmark data set and on implemented models within a simulation framework. The empirical baseline of this work is a highly accurate benchmark data set of 2674 human trajectories measured under real life conditions in a bidirectional corridor with a novel data collection approach using the low-cost sensor Microsoft Kinect. Our innovative human detection and tracking algorithm is based on agglomerative clustering of privacypreserving depth data captured from an elevated view with multiple Kinects providing a Pedestrian Detection Rate of up to 94% and a Multiple Object Tracking Precision of 4 cm. The proposed simulation framework is built on a scalable and flexible system architecture to easily integrate different simulation models. Hence it allows for consistent and efficient model calibration and validation. Three approaches of the Social Force model, a Cellular Automaton model, the Optimal Reciprocal Collision Avoidance model and two variants of the Optimal Steps Model were implemented in the simulation framework. As a first step towards model verification and validation, we have simulated selected test cases from the RiMEA-Guideline, which aspires to define a minimum standard for evacuation analysis. The potential of the simulation framework for easy and quick switching and combining models is demonstrated using two real world case studies: first, two versions of the Social Force approach and the Optimal Reciprocal Collision Avoidance model are applied within our framework for passenger flow simulation during boarding and alighting of a train. Second, the ability of the proposed simulation framework to Combine modeling approaches with varying granularity is demonstrated for high volume passenger flows in a subway station. In the evaluation, the predictions of seven models implemented within the simulation framework are compared to the empirical benchmark data from the Kinect-based tracking and among each other. In order to establish a solid baseline for model comparison, every model is calibrated first on a subset of the benchmark data. Two methodologies for calibrating Social Force based models on the individual trajectory level are presented, i.e. model estimation by nonlinear least square methods and comparison of real and simulated trajectories. Furthermore, we introduce a structured evaluation environment based on measures to assess individual model capabilities of representing important microscopic and macroscopic characteristics of human movement behavior. A simulation-based calibration procedure is applied in our simulation framework to estimate the parameter values for the different modeling approaches with the defined set of evaluation measures. It was found that the calibration has improved the fit to the observed data set in all models. However,the grade to which individual models can be influenced by the calibration varies. The evaluation also revealed that the investigated models have diverse capabilities concerning transferability to an independent data set. Our presented evaluation technique can easily be applied to a wider range of pedestrian modeling approaches by including them as separate, additional modules in the Simulation framework. For future studies this will enhance the understanding of individual model characteristics and the comparison of novel modeling approaches to existing ones.