Decoding micro-social interactions in public space: a computer-vision-based method

Abstract As cities strive for greater liveability, data-driven methods can enhance our understanding of public space behaviours and social interactions. Recent developments in computer vision technologies have significantly advanced the accuracy of micro-scale human behaviour detection, but there is a lack of methodologies that capture relational, nuanced behaviours within specific spatial and temporal environments. This paper presents the development of a computer vision and machine learning-based methodology to analyse co-presence and micro-social interactions in urban spaces, introducing new metrics for spatial behavioural analysis. The methodology was tested on a 22.5-min video dataset obtained at a university campus, demonstrating its capacity for trajectory analysis and detecting nuanced interpersonal behaviours including encountering, congregating, approaching and avoiding. Human observers validated the computer-generated behaviour classifications, achieving high agreement levels and demonstrating the system's accuracy in detecting diverse pedestrian interactions. The approach successfully offers fine-grained analysis of social behaviours and spatial patterns of co-presence, revealing how urban morphology influences social interaction hotspots. It advances environment-behaviour research by providing scalable, automated tools for detailed, data-driven analysis of public space vitality, with potential applications in urban design, social sciences, and policy-making.