Designing with Human Behavior: A Researchby-Design Exploration of AI-Assisted Pose-Based Activity Analysis for Small-Scale Spatial Optimization in Architecture

Description

This research investigates the integration of Artificial Intelligence (AI)-based pose estimation into Post-Occupancy Evaluation (POE) to facilitate data-driven spatial optimisation in small-scale architectural environments. Traditional POE methods involve observers manually recording human behaviour, which often fail to capture real-time, non-verbal behavioural nuances (Neto et al. 2020; Duroosaiy et al., 2019; Roberts & Edwards, 2022). This study addresses this gap by proposing a digital workflow that bridges human behavioural data with spatial optimisation design. The research developed an AI software that utilises the YOLOv7-pose (M et al. 2023; Pires et al. 2025) framework to extract skeletal key points from video inputs in real-time. This data is processed to generate spatialized heatmaps that categorise occupancy duration and specific activities, including standing, sitting, walking, and laying. The methodology followed a two-stage validation process:

1)Technical Feasibility: Initial testing at a sunken plaza (Site 1) to calibrate the computer vision algorithm.
2)Practical Application: A case study on Fengmen Heng Street (Site 2), a historic market street in Suzhou, characterised by high pedestrian density and a deficiency in public rest facilities.

The heatmap analysis revealed significant spatial underutilization and a “mismatch” between existing infrastructure and actual resting behaviours. These behavioural insights directly informed targeted design interventions. The comparative analysis of the two sites demonstrates that while Site 1 validates the technical robustness of the AI tool, Site 2 confirms its efficacy as a diagnostic instrument in complex, real-world urban contexts. The study identifies a critical limitation: the impact of low-resolution video inputs on pose estimation accuracy. Furthermore, the findings emphasise that the AI system functions as an analytical assistant rather than an autonomous generator; human designers remain essential for interpreting contextual and cultural nuances to translate heatmaps into qualitative design solutions. This research contributes a validated “research-by-design” workflow. It demonstrates that the synergy between AI-assisted pose analysis and professional architectural judgment provides a more granular, evidence-based approach to small-scale spatial optimisation, moving beyond conventional observational techniques.

Period	1 Jan 2026 → 30 Jul 2026
Examination held at	Design School
Degree of Recognition	Local