Deep learning-based behavioral pattern extraction from transit station videos modeling structural gender differences using CNN and Vision Transformers

Abstract

Urban transit stations are among the most behaviorally complex environments in public transport systems, yet planners still lack scalable methods for converting continuous video observations into interpretable, planning-relevant behavioral evidence. Existing transportation research relies heavily on surveys and self-reports that capture perceptions rather than observed behavior, while computer-vision studies achieve high detection and tracking accuracy but rarely transform trajectories into standardized indicators usable for station design, capacity assessment, or equity analysis. This study addresses this methodological gap by developing and demonstrating an interpretable behavioral measurement framework that extracts trajectory-based indicators waiting duration, interpersonal distance, trajectory entropy, walking speed, and congestion exposure from transit-station video and quantifies context-dependent differences associated with visually inferred gender presentation. A hybrid Convolutional Neural Network–Vision Transformer (CNN–ViT) architecture is employed to produce reliable, identity-preserving trajectories under the occlusion, density, and illumination conditions typical of real stations. The hybrid design combines the local feature extraction strengths of CNNs with the long-range contextual modeling of Vision Transformers, yielding higher-quality trajectories than conventional single-architecture baselines. Explainable AI techniques are used solely to enhance transparency of the computational pipeline, not to validate theoretical constructs. Behavioral differences are interpreted strictly as patterns associated with visually inferred gender under specific environmental conditions, maintaining a clear distinction from lived-experience constructs of gendered mobility. The framework bridges the disconnect between algorithmic outputs and the analytical needs of transportation planning by delivering compact, reproducible indicators that can inform passenger-flow management, station design evaluation, and operational analysis while respecting the evidential limits of observational data. The principal contribution is therefore methodological: an evidence-based measurement bridge that strengthens, rather than replaces, existing transportation knowledge.