Global Manufacturing Campus Intelligent Safety Environment
A flexible intelligent safety response system changing the paradigm of safety control for ultra-large sites.
Building a Flexible Safety Response System
The Global Manufacturing Campus has introduced EVA to secure worker safety and respond quickly to hazardous situations.
EVA x Global Smart Camera: Intelligent Safety Environment
We built an intelligent safety control environment by combining global partner's smart camera infrastructure with EVA.
Safety managers can add cameras effortlessly and write natural language detection scenarios without complex settings or development.
This ensures that even as the site changes, the system can quickly reflect those changes to maintain real-time response capabilities.
A Flexible Safety System Focused on Real-time Response
Key Detection Targets
We constantly monitor for PPE compliance (masks, helmets) based on hazardous material handling standards.
We also detect abnormal situations such as collapses or long-term neglect that can occur in 24-hour shift environments and solo workspaces.
Major Activity Cases
In the early stages of EVA deployment, various false positives can occur due to the complexity of the actual work environment.
However, when on-site managers provide explanations and feedback, EVA gradually refines its judgment criteria based on them.
In this process, detection accuracy continues to improve.
Below, we introduce three representative activities that led to these improvements.
Activity 1: Refining Detection Scenarios
| Case | Stage 1 (Initial) | Stage 2 (False Positive Description) | Stage 3 (Enrich Agent Structuring) |
|---|---|---|---|
| Fall Detection | Let me know if a person falls. | Notify me if a person collapses on the floor, excluding those sitting in chairs or standing. | Current Case Detecting a person collapsed on the floor Detection Steps - Person exists - At least one person is collapsed on the floor Exceptions - Person sitting or standing - Difficult to identify the body - Hard to distinguish from lying down - Body obscured more than 50% - Obscured by floor objects or structures |
| Mask Non-compliance | Notify me if someone is not wearing a mask. | Among people working in chairs, let me know who is not wearing a mask. Exclude people using laptops/phones. | Current Case Detecting mask non-compliance among workers sitting in chairs Detection Steps - Sitting in a chair - Working (determining behavior) - Not wearing a mask on the face Exceptions - Difficult to judge work activity - Face or worn objects obscured |
Simple lists of conditions often fail to reflect the complexity of actual sites.
EVA structures user descriptions and exceptions step-by-step,
enabling the AI to understand not just 'what to see' but also 'what not to see.'
The Enrich Agent performing this role is a core component for scenario completeness.
Activity 1: Feedback on False Alarms
Feedback provided by users during false positives is stored in VectorDB,
preventing identical alarms from repeating in similar situations.
Through over 250 cumulative feedbacks, EVA learns on-site characteristics and continuously refines judgment criteria.
This feedback loop is a key mechanism that increases system reliability as it operates.
Activity 2: Foundation Model Updates
EVA is not tied to a specific model;
it can flexibly combine various Vision Models or transition to the latest LLM/VLM models according to site characteristics and operational needs.
In this case, we used RT-DETR V2 to quickly detect person objects,
and combined it with VitPose to verify actual posture, significantly reducing false positives.
Technical details can be found in our Tech Blog.
Furthermore, by validating various newly released VLM models and updating to Qwen3-VL 8B, the overall detection performance was improved,
and processing speed was increased by 50% through a lighter model configuration.
Activity 3: Area-based Detection
Area-based detection has been added since EVA v2.3.0.
Users can specify areas of interest so EVA focuses only on those zones,
effectively reducing false positives caused by unnecessary actions outside work areas.
EVA x Workflow: From Detection to Action
EVA's detection results are combined with the Workflow Builder to expand into actual operational actions.
Integrating with Power Automate and n8n allows for automatic response flows based on detection severity,
with results provided in analysis reports to support manager decision-making.
EVA x Workflow: Physical Action Integration
In cases of detected high-risk scenarios,
Workflows can immediately control physical devices on-site, such as sirens or emergency lights.
Through this, EVA evolves beyond a simple analysis tool into an action-oriented safety platform that drives real field behavior.
Details can be found in our Tech Blog.
Efficient Scalability
With simple integration with global partner's smart camera infrastructure, you can simultaneously secure cloud-based scalability and EVA's flexible detection operations.
This provides a foundation for expanding intelligent safety environments in a cost-effective and efficient manner, even for large-scale sites.
