Learning and Simulating Human Movement in Public Spaces

Abstract

Prior urban studies initiatives have used observational methods to capture and model human movement in public spaces, resulting in static visualizations that highlight generalizable behaviors and trends. Motivated by the limitations of static representations to portray dynamic movement, this thesis presents a proof-of-concept tool that uses a data-driven approach to learn and simulate human movement in urban public spaces. Human movement is learned with a reinforcement learning model proposed by Kitani et al. in “Activity Forecasting” (2012). This model looks at human interactions with the static features of a scene -- buildings, cars, grass, etc. -- to predict paths of movement. I retrain this model with videos of pedestrian-friendly scenes to get a set of feature weights that convey the influence of each static feature on human movement. With these feature weights, I feed sample images of public spaces into an Optimal Control (OC) model that forecasts a trajectory between a specified source and destination. The user also has the ability to paint over the image to add in additional static features to see how the predicted trajectories change accordingly. I use an adapted form of Dijkstra’s Shortest Path algorithm (DSP) to find the maximum likelihood single-line path from source to destination. I stitch motion capture figures from CMU's MoCap dataset along the path to simulate this movement. My findings reveal that retraining the OC model with more pedestrian-friendly scenes improves the model’s sensitivity to static features like “grass” and “pavement” while decreasing its sensitivity to features like “car.” Further, an analysis of my path-finding algorithm shows that the overall cost of the paths outputted by DSP is lower than that of straight-line paths, highlighting the plausibility of the simulated patterns of movement. This tool thus presents an al- ternate, dynamic form of spatial representation for architects and urban planners, having the potential to increase understanding of human interaction with spaces that could lead to increasingly people-sensitive built environments.