Data-driven pedestrian simulation
Date of Issue2017-03-29
School of Computer Science and Engineering
Parallel and Distributed Computing Centre
Over the past few years, crowd simulation has been an active research field with an increasing attention from different research areas such as virtual environment, object tracking, computer animation and civil planning. These applications require the synthesized pedestrians to move in a realistic and believable manner, in order to have an excellent user experience or to draw some reliable conclusions based on the simulation results. Various crowd simulation models have been proposed over the years, from the earliest macroscopic models inspired by fluid dynamic to rule-based and force-based models where proper mechanisms are designed based on model designer's domain knowledge. More recently, real world pedestrian motion data has been introduced in modeling crowd behavior. These models are referred to as data-driven models in this thesis. Some researchers calibrate the parameters in their models using human motion data to improve the accuracy of motion prediction. Other data-driven models extract motion behavior from human motion data and then utilize the learned knowledge to drive pedestrians' behaviors in real time. This kind of models are referred to as example-based models in this thesis. Example-based models are more likely to produce realistic crowd motion simulation compared to other models, because motion data extracted from real pedestrian preserve the variety and complexity of crowd motion in real world. However, the existing example-based models fail to reproduce crowd motion both accurately and efficiently, at the same time, during the simulation. Besides, pedestrians behave quite differently in high density scenarios from those in low density cases. Collective motion behaviors such as lane formation are commonly observed in crowded scenario while in low density scenarios pedestrians generally have more freedom in avoiding collisions. To address the above issues, this thesis proposes two crowd models focusing on simulating collision avoidance and collective motion behavior respectively. A clustering-based model is proposed to reproduce pedestrians collision avoidance behavior by introducing a more accurate method to extract collision avoidance behaviors from human motion data and a more accurate and efficient approach to query the extracted examples in real time. Secondly, a role-dependent model is used to generate realistic motion behavior in crowded scenarios based on the observation that pedestrians either behave as a leader to lead the way or as a follower that follows behind others in front. The decision as to whether a person is acting as a leader and or as a follower is not static, and can change from time to time, dynamically, based on sensing the immediate crowd conditions in the direction that the person is heading. The main evaluation method applied by most crowd models so far is to visually compare the trajectories simulated by different models. This subjective approach can give a direct impression of the believability of the simulated trajectories but fails to provide a thorough and comprehensive evaluation of crowd models. Although some existing research does compare their proposed models with the existing models by using real-world crowd videos as the ground truth to demonstrate their models' usefulness, much of the studies consider only some of the existing models and compare them using limited number of quantitative metrics. Therefore, in this thesis, a detailed quantitative evaluation of the proposed crowd models, using the real-world crowd motion data with comprehensive metric measures, is described. A comprehensive study of the proposed models with the existing ones is also represented.