Key Takeaways
- The Passenger Movement Estimation System (PMES) uses AI‑driven analysis of a single CCTV frame to predict passenger trajectories and prevent subway door entrapment before it occurs.
- Empirical validation shows that 97.85 % of passengers descending station stairs move directly toward train doors when a train is present or approaching.
- Real‑time classification of passenger movement (ascending, descending, passing) achieves 97.58 % accuracy with an object‑detection model.
- The ultra‑lightweight SD Net (Subway Door Network) enables deployment on limited‑computing hardware, making the solution practical for existing subway infrastructures.
- A Decision Support System (DSS) provides train operators with optimal door‑closing timing recommendations and passenger‑warning guidelines, enhancing safety while reducing delays.
- The study results from a global, interdisciplinary collaboration among Sungkyunkwan University, KAIST, and Texas Tech University, and are slated for publication in IEEE Transactions on Intelligent Transportation Systems (top 1.89 % in JCR).
Overview of the Research Initiative
The research team, led by Professor Jo Woon Chong from Sungkyunkwan University’s School of Electronic and Electrical Engineering, partnered with experts from KAIST and Texas Tech University to develop the Passenger Movement Estimation System (PMES). PMES leverages artificial intelligence to analyse closed‑circuit television (CCTV) footage and forecast passenger movements on station platforms. By predicting where passengers will be relative to train doors, the system aims to avert door‑entrapment incidents before they happen, shifting safety strategies from reactive to proactive. This approach addresses a persistent safety challenge in subway networks, especially those equipped with platform screen doors (PSDs) or automated train control systems.
Behavioral Insights from the Passenger Trajectory Model
A core component of the study is the Passenger Trajectory Model (PTM), which was empirically validated using extensive video data collected at multiple subway stations. The PTM revealed that, when a train is either stationed at the platform or approaching, 97.85 % of passengers who are descending the stairs proceed straight toward the train doors. This highly directional behavior provides a reliable basis for forecasting risk zones. The research team highlighted that recognizing this pattern allows the system to anticipate potential entrapment scenarios well before passengers reach the dangerous proximity of the doors.
Real‑Time Movement Classification Accuracy
To operationalize the PTM, the team classified passenger motion on station stairs into three mutually exclusive categories: ascending, descending, and passing (lateral movement across the stairwell). Using a state‑of‑the‑art object‑detection deep‑learning network, they achieved real‑time classification with an impressive accuracy of 97.58 %. This high precision ensures that the system can reliably distinguish between passengers who are moving toward the train (and thus at risk) and those who are merely traversing the stairwell for other purposes, thereby minimizing false alarms.
Design of the Ultra‑Lightweight SD Net
Recognizing that many subway stations operate with constrained computational resources, the researchers introduced SD Net (Subway Door Network), a custom‑built, ultra‑lightweight neural network architecture. SD Net retains the detection accuracy required for reliable movement classification while dramatically reducing parameter count and computational load. As a result, the model can run efficiently on modest hardware such as edge‑computing devices or existing CCTV processing units, facilitating widespread deployment without costly infrastructure upgrades.
Integration with a Decision Support System
Beyond detection, the team developed a Decision Support System (DSS) that translates movement predictions into actionable guidance for train operators. The DSS recommends optimal door‑closing timestamps based on the real‑time flow of passengers toward the doors, thereby balancing safety with operational efficiency. In tandem, the DSS generates passenger‑warning alerts—such as auditory signals or visual cues on platform displays—when an imminent entrapment risk is detected. This dual‑layered approach equips both staff and riders with timely information to prevent accidents.
Interdisciplinary Collaboration and Global Impact
The project exemplifies the power of interdisciplinary cooperation. Professor Lisa Lim from KAIST contributed expertise in civil and environmental engineering, particularly regarding platform layout and passenger flow dynamics, while Researcher Yifan Li from Texas Tech University brought deep knowledge of electrical and computer engineering, focusing on embedded AI implementation. Their combined insights ensured that the PMES solution is both technically robust and practically suited to real‑world subway environments. The study’s findings are set to appear in IEEE Transactions on Intelligent Transportation Systems, a journal ranked in the top 1.89 % by Journal Citation Reports, underscoring its scholarly significance.
Implications for Subway Safety and Operational Efficiency
By converting passive CCTV surveillance into an active safety net, PMES offers a pathway to dramatically reduce door‑entrapment accidents, which continue to cause injuries and service disruptions worldwide. Early risk detection enables operators to adjust door‑closing protocols preemptively, decreasing the likelihood of injuries and minimizing unnecessary train delays caused by emergency stops or door‑reopen cycles. Furthermore, the system’s lightweight nature ensures that even older subway lines with limited processing capabilities can adopt the technology, promoting equitable safety improvements across diverse transit networks.
Future Directions and Scalability
Looking ahead, the research team envisions extending PMES to incorporate multimodal data sources, such as infrared sensors or wearable device signals, to enhance prediction robustness under varying lighting and crowd conditions. They also plan to explore adaptive learning mechanisms that allow the model to continually refine its understanding of station‑specific passenger behaviors. As more cities adopt platform screen doors and automated train control, the proactive safety framework established by this study could become a standard component of intelligent transportation systems, fostering safer, more reliable urban mobility on a global scale.