Agent-agent interaction as a component of agent-environment interaction in the modelling and analysis of pedestrian visual behaviour

This multidisciplinary project involves concepts from architectural design, statistical learning, machine vision, and human ecology. The focus is on analysing how pedestrians’ dynamic behaviour in space is influenced by the environmental design of different architectural scenarios. This paper presents a multi-agent pedestrian simulation and analysis system that supports agent-to-agent interactions, different spatial desires, and interpersonal distance. The system simulates and analyses pedestrian spatial behaviour with combined focus on movement trajectories, walking speed, and the visual gaze vector. The analysis component relies on learning a statistical model characterising normal/abnormal behaviour, based on sample observations of regular pedestrian movements without/with the impacts of significant visual attractions in the environment. Using the example of Wheeler Place in Newcastle (Australia) our pilot experiments demonstrate how pedestrian behaviour characteristics can depend on selected features in the visual environment. The presented system will allow architects and urban designers to obtain better assessment of planned urban spaces and streetscape characteristics and their impacts on pedestrian behaviour