Crowd Analysis in Subway Platforms via Semantic Segmentation and Perspective Compensation

ABSTRACT

Passenger congestion in subway station waiting areas has emerged as a critical challenge in modern urban transit systems, particularly under high‐density conditions. Conventional object detection methods such as You Only Look Once (YOLO) often struggle in visually complex environments due to overlapping objects, occlusion and limited spatial resolution for precise object boundaries. To overcome these limitations, a semantic segmentation (SS)‐based framework is proposed for crowd congestion level classification, designed to estimate pixel‐level occupancy in closed‐circuit television (CCTV) imagery without relying on bounding boxes. Unlike conventional density estimation approaches that primarily focus on passenger counting, the proposed framework quantifies crowd conditions using occupancy ratios derived from pixel‐level segmentation results and classifies congestion levels based on space utilisation characteristics. Two additional correction models, U‐ASD Net and perspective‐aware convolutional neural network (PACNN), were incorporated to compensate for perspective distortion caused by varying distances and crowd densities within the scene, which result in significant variation in object scales across foreground and background regions. A dataset comprising 72,000 images per station from two major dense traffic transfer stations was analysed, and occupancy ratios were clustered into six levels of service (LOS) using a Gaussian Mixture Model (GMM). The SS‐based method demonstrated improved classification performance compared to the object detection baseline. Although correction‐integrated models showed slightly reduced average accuracy, their performance improved in severe congestion scenarios, indicating enhanced stability under visually degraded conditions. Statistical analysis confirmed significant model‐level variation, supporting the potential of segmentation and adaptive correction for practical application in transit surveillance environments. The proposed framework can support real‐time crowd monitoring and congestion management by providing quantitative occupancy‐based indicators for transit facility operations. Future work will focus on expanding the training dataset and improving model generalisation across diverse environments.