<p>Metro systems face challenges in managing localized congestion and ensuring passenger safety, particularly under high-density passenger flows and unexpected disruption scenarios. Conventional crowd detection methods often fail due to severe occlusion, insufficient coverage, and excessive latency, making them unreliable for providing operators with accurate risk identification and control support. To address this fundamental problem, this paper proposes a vision-driven framework for risk-point identification tailored to the complex spatial structure of metro stations. The framework adopts a world-coordinate-first multi-camera fusion strategy combined with CAD (Computer-Aided Design) floor-plan projection to achieve station-wide coverage and spatial interpretability; it introduces a density-speed dual-factor risk index to overcome the limitations of single-density indicators in capturing dynamic anomalies; and it incorporates a baseline-increment mechanism to effectively distinguish recurrent congestion from emerging anomalies, supporting fine-grained risk management. Empirical studies at two representative transfer hubs in Shanghai—Station A and Station B—demonstrate that the framework consistently achieves recall rates above 0.9 and 100% coverage of risk points in both routine and disruption scenarios, while sustaining real-time performance at 28 FPS. These results confirm that the proposed framework effectively bridges the gap between macroscopic passenger flow forecasting and microscopic risk sensing, offering a promising technical foundation for pilot-stage operational safety management and fine-grained risk sensing.</p>