Modular and Composable AI Video Analytics for Construction Safety
Construction Site Safety
Challenge Statement Owner
Leighton Asia, a member of the CIMIC Group, delivers a portfolio of high-profile infrastructure projects and offers a comprehensive range of services spanning key sectors including civil infrastructure, advanced technology, buildings, energy and resources, and MEP for over 50 years across Asia. In Singapore, the company’s projects include the Deep Tunnel Sewerage System Phase 2, North South Corridor Project, Choa Chu Kang Waterworks, and a data centre project.
Background
There are challenges in deploying AI video analytics (VA) for construction safety due to the complexity of high-risk activities such as working at heights, operating near machinery, and performing lifting operations. The construction industry as a whole is inundated with alerts from current VA systems, resulting in alert fatigue where teams cannot effectively respond to the sheer volume of notifications.
While over 90% of detections were technically correct, only 10-20% were actionable. Around 70-80% of alerts lacked context, creating alert fatigue as safety teams struggled to respond to the sheer volume of irrelevant notifications.
Existing solutions also fail to adapt to dynamic site conditions, such as shifting barriers, cranes, and machinery zones, resulting in false positives. Improving AI video analytics now can significantly enhance construction safety and efficiency by reducing alert fatigue, ensuring focus is on meaningful alerts only.
Current VA systems detect compliance breaches, for example, if a worker is near an open edge, even when that worker is wearing a harness and is thus safe. AI is technically correct most of the time, but lacks contextual interpretation. Too many non-actionable alerts reduce impact and confidence for site teams.
Having adequate confidence in the alerts is critical for the future success of VA. For example, in order to maintain a robust process, our site teams have implemented a procedure that enables timely mobilisation to resolve an alert. We utilize ‘Telegram’ to receive real-time notification alerts, then each alert is assigned to the respective person in charge, who is then responsible for reviewing the nature of the alerts and providing a suitable and timely response. All responses are monitored and recorded, and the data is displayed on a centralized dashboard for tracking and accountability. This further highlights the need to improve the system outputs and reduce wasted efforts.
The opportunity lies in developing a modular and composable VA solution that allows end users to combine use cases or create their own use cases, to effectively address high-risk activities such as working at heights and near machinery. Such a solution would offer site-specific analytics and alert customisation.
Current safety standards are rigid (distance-based), but AI has the potential to introduce dynamic conditions. The industry could benefit from objective AI-based assessments, improving how safety is measured and managed on site.
The Challenge
How might we reduce alert fatigue from AI video analytics in construction by developing a modular and composable framework that combines multiple conditions to deliver relevant, contextual, and actionable alerts tailored to high-risk activities?
Requirements
Functional Requirements
- Shift from isolated safety scenarios to a composable VA system that empowers users to combine detection logic tailored to site-specific needs
- Combine multiple high-risk conditions into contextual alerts, such as (Leighton Asia’s priorities):
- “Open edge + No harness”
- “Near machinery + No Access”
- “Under lifted loads + No Access”
- "Smoke or fire + No access in isolated or unmanned areas"
- Adapt to dynamic site layouts, recognising barriers and zones in real time
- Enable faster configuration and reuse of detection workflows across projects without rebuilding from scratch
Technical Requirements
Core Requirements:
- Provide a modular workflow builder with user-friendly, drag-and-drop interface for easy updates and configuration
- Support multi-condition logic (more than 2 conditions in future) with scalable design
- Integrate with existing CCTV infrastructure (e.g., 2MP cameras, AI edge hardware, etc.) for cost-effectiveness
- Incorporate a contextual interpretation layer to filter out technically correct but non-actionable alerts
Additional Preferences:
- Enable portability of detection workflows across projects
Note: Given variability in building tolerances and alignment needs, human oversight in final positioning is acceptable.
Expected Outcomes
- Achieve over 85% accuracy in use case combinations (2 or more conditions)
- Reduction of alerts that can be considered compliant (ie. alerts detected that are correct based AI’s standard parameters but is not considered non-compliance due to operational constraints that can be observed visually)
- Improvement in safety outcomes and team confidence by ensuring attention is directed to critical alerts
- Acceleration of deployment cycles by reusing detection workflows across projects
Deployment Environment and Constraints
The system must operate effectively in dynamic construction environments, handling changing layouts, barriers, and machinery zones. It must support use in Singapore initially, with the ability to extend to Hong Kong and Malaysia.
Proof-of-concept (POC)/Pilot Support
Leighton Asia will collaborate with innovators to develop a POC at their ongoing data centre project in Singapore, where cameras are already deployed and current modules are running. The POC will focus on the above mentioned conditions:
- “Open edge + No harness”
- “Near machinery + No Access”
- “Under lifted loads + No Access”
- "Smoke or fire + No access in isolated or unmanned areas"
POC validation will prioritise detection accuracy, alert relevance, and usability.
Leighton Asia will provide:
- A testing site in Singapore
- Deployment assistance and installation support
- Collaboration with digital and innovation teams
- Data capture and validation for solution development
Commercialisation and Scaling
If effective, Leighton Asia will commit resources to implement the solution across its projects in Singapore, and later extend to sites in Hong Kong and Malaysia. With proven results, the solution has potential for broader adoption across the construction sector, as the issue of alert fatigue is an industry-wide challenge.
