AI pedestrian detection system for railway crossings

Project summary

Railway level crossings represent critical safety vulnerabilities, as pedestrians and vehicles interact with train traffic. Traditional monitoring systems rely heavily on manual observation or basic motion detection that cannot reliably differentiate between different types of object. RSK Business Solutions developed and piloted a computer vision AI system to address these challenges, demonstrating how artificial intelligence can reduce false alerts, improve pedestrian detection and support safer railway operations without requiring costly infrastructure overhauls.

The background

According to the Office of Rail and Road (ORR), there were hundreds of near misses with pedestrians at level crossings in the UK during each year, with five fatalities reported in 2022/2023 (ORR annual statistical release, 2023). The Rail Accident Investigation Branch (RAIB) has consistently highlighted human error and risk-taking behaviour as contributing factors. Despite more than £100 million of Network Rail investment since 2010, existing safety systems remain heavily reliant on human monitoring or basic detection technology.

This research project led by RSK Business Solutions was initiated to explore how AI-powered computer vision could enhance detection accuracy, reduce operator workload and deliver scalable improvements to pedestrian safety at level crossings.

The challenges

Currently, there is heavy reliance on manual observation for level crossing operation, which is resource-intensive and can be inconsistent. The conventional motion detection systems that are in use are unable to differentiate between pedestrians, vehicles and other objects, resulting in high false-positive rates. In addition to this, existing CCTV systems have limited capability of providing real-time, intelligent detection, and there is a high cost associated with upgrading infrastructure, making network-wide improvements difficult.

The solutions

To address these challenges, RSK Business Solutions developed a computer vision AI system in consultation with industry safety officers, using a structured four-stage process that utilised existing CCTV infrastructure across the network and had minimal additional hardware requirements.

The four stages were as follows.

• Research and analysis: This involved a review of existing detection systems and their limitations; analysis of Office for Rail and Road and Rail Accident Investigation Branch incident reports; and stakeholder workshops with safety officers, control room operators and risk management teams.
• Model development: Over 42,000 images of railway crossings were collected and annotated under varied conditions. These were then used to train multiple model variants using the ‘you only look once’ (YOLO) object detection framework, integrated with Roboflow for dataset management and optimisation. Spatial-temporal algorithms to detect pedestrians entering danger zones were then developed from these data.
• Testing and refinement: The model underwent laboratory testing with recorded CCTV footage and controlled field trials at a non-operational test facility. A live pilot deployment at three high-risk crossing locations in southern England was then carried out and iterative model updates based on operator feedback and performance analysis were applied.
• Deployment and monitoring: Following successful testing, the computer vision AI tool was integrated with existing CCTV infrastructure and control systems at pilot sites. The installation required minimal additional hardware, reducing implementation costs, and training was provided to operators and control room staff. Monitoring protocols for continuous performance improvement were established for these pilot sites.

The impact

Pilot testing demonstrated the strong potential of AI-powered pedestrian detection by reducing false positives compared to motion detection, enabling operators to focus on genuine risks. It was effective in the early detection of pedestrian violations, providing operators with additional time to respond.

Indications of reduced near-miss incidents at trial sites aligned with the Office for Road and Rail’s CP7 target of 50% risk reduction. As a result, operator efficiency was improved, with monitoring hours reduced during peak periods; there was increased situational awareness through intelligent visual alerts and automated incident logging; and a reduction in on-site operator exposure to hazards, with fewer manual surveys required.

Implementing the technology offers potential cost savings through reuse of existing CCTV infrastructure. The project demonstrates how AI can transform level crossing safety effectively and efficiently. With further development and larger-scale deployment, the system has the potential to deliver measurable safety gains, operational efficiencies and cost savings across the UK rail network.