Representative Case Study: On-Premise PPE Detection for a Multi-Zone Industrial Site

On-Premise PPE Detection for Industrial Site

The operational challenge

A growing industrial site wanted a proactive way to identify missing PPE in high-risk work areas without sending all camera footage to a cloud service.

CCTV available was utilized for the purpose of post-analysis but was not able to give timely and systematic safety events.

The customer needs local processing, support for a variety of IP cameras, workflow of audit events, and specific design that can be applied even under adverse conditions like lack of light, unstable network, and in situations in which people hide behind the gear.

The required PPE varied by zone. In a loading area, the focus was hard hats and hi-vis vests. In a fabrication area, protective eyewear and face protection could also be required.

The key requirement was not “detect PPE somewhere in a frame”; it was to identify a person entering a defined zone without the PPE required for that zone and notify the right safety team without creating alert fatigue.

Solution architecture

We designed the deployment design on the NVIDIA Jetson Orin Nano Super Developer Kit, chosen for its small edge-AI design and advanced inference ability for vision applications. The Jetson was operated by Ubuntu obtained from NVIDIA JetPack, enabled by NVIDIA drivers, CUDA libraries, multimedia accelerators, TensorRT, and supported by containers.

Existing ONVIF-compatible IP cameras delivered H.264/H.265 RTSP streams over a segmented local network. GStreamer and DeepStream components were used to ingest, decode, batch, and process the streams efficiently.

The inference service performed person and PPE detection, while a tracking layer assigned a temporary identity to each visible person across successive frames. This reduced repeated alerts when the same worker remained in view.

A rule engine joined three sources of information: detected objects, configured polygon zones, and the PPE policy for each zone. The event logic checked whether a tracked person was inside a zone long enough to be considered present and whether the required PPE was visible with sufficient confidence.

A configurable grace period and persistence rule prevented an alert from being raised when a helmet was briefly hidden by a raised arm, a pallet, or another worker.

The system separated video analysis from reporting. The live video feed was processed locally. Once an event occurred, the system saved a timestamped image, camera ID, PPE rule outcome, and the respective confidence value in the internal database.

The event data was accessed from the dashboard by an internal API.

AI model and data engineering

A generic PPE model is rarely enough for a production deployment. Site-issued helmets may have unusual colours, vests may be covered by jackets, face shields may look different from safety glasses, and camera angles may vary sharply. We therefore planned a data pipeline using representative footage collected under approved site procedures.

The annotation schema included person, helmet, high-visibility vest, safety glasses, face shield, glove, and other agreed classes.

Multiple samples were chosen from various shifts, weather conditions, camera perspectives, crowd population, body postures, blockages, and distance ranges. The training, evaluation, and test sets were separated, preventing exaggerated results obtained from almost identical specimens.

Once trained and validated, the model was passed to ONNX and later turned into a TensorRT engine relevant to the Jetson platform.

The end-to-end performance profiling was done instead of recording model inference performance only. The good acceptance criteria are the number of cameras used, resolution of the source, frame rate, GPU/CPU usage, thermal indicators, memory consumption, event timing, and stream recovery performance.

All these things are important as a load on the decoder, overlays for analytics, writes to database, and instability of the network can become a factor in the production process.

Privacy-aware event reporting

The organization believed that having an overly-safe workflow would not fit in with their ideals. We developed a reporting practice that focused on incidents labeled as safe.

The platform interface contained helpful information including a smaller version of the image, details regarding the date and time of the event, location, details about PPE use, event state, and information introduced by the supervisor. Whenever possible, the stored photo was subject to face-blurring.

Access to the system was supposed to be given based on users’ roles which meant that system administration, safety reviewers, and read-only managers had only the access they needed for their work.

Retention schedules could automatically delete event images after the organisation’s approved period, while preserving anonymised aggregate metrics such as events by zone, PPE category, time period, and resolution status.

By using this method, health-and-safety teams can obtain usable information without treating each movement of their employees as a record that requires reporting.

It helps to have a straightforward discussion about governance before implementing the system: which locations will be included, who will analyze notifications, how workers will receive information on new events, how long the evidence will be kept, and how disagreements regarding detecting certain actions will be solved.

Validation and deployment

Before production rollout, we validate the system in stages. First, recorded streams confirm camera connectivity, model loading, and event rules. Next, controlled site walkthroughs test expected PPE and deliberate non-compliance scenarios. Finally, a monitored pilot measures real operational performance and tunes thresholds, zone boundaries, grace periods, and notification routing.

We include negative scenarios in the test plan: camera disconnect, RTSP reconnect, Jetson restart, database unavailability, low-light periods, congested scenes, PPE partially outside the frame, and dashboard-network loss. Alerts should be queued or safely retried rather than silently lost. Device-health signals report temperature, storage, CPU/GPU utilisation, stream status, and service availability so the safety team knows when the system is not in a reliable monitoring state.

Jetson Orin Nano Super PPE Detection Development Services

Adequate Infosoft provides end-to-end NVIDIA Jetson Orin Nano Super development services for organisations that need local, responsive, and controllable PPE monitoring. We begin with a practical discovery workshop covering which zones matter, what PPE rules apply to each zone, how many cameras are needed, what constitutes a reportable event, and who may access footage or event metadata.

This approach helps prevent a common failure mode in computer-vision projects: selecting an AI model before defining the operational workflow.

Industrial Video Architecture and Camera Integration

Our engineering team designs the video architecture around the actual deployment site rather than assuming every IP camera behaves identically. We assess camera position, lens field of view, illumination, mounting height, RTSP stream stability, network segmentation, and the expected size of a person in each frame.

Industrial environments can introduce challenges such as:

  • Small helmets or safety equipment at long camera distances
  • Heavy backlighting around entrances
  • Reflective PPE
  • Forklifts or machinery temporarily blocking workers
  • Workers moving between indoor and outdoor environments
  • Variable lighting throughout the day
  • Unstable or high-latency video streams

When required, we recommend separate detection zones instead of attempting to make one camera cover an unrealistic area.

GPU-Accelerated Edge AI Pipeline

On the Jetson device, we build a GPU-accelerated computer-vision pipeline using NVIDIA JetPack, CUDA, TensorRT, GStreamer, and DeepStream, where appropriate.

The pipeline can ingest:

  • RTSP camera streams
  • USB cameras
  • CSI cameras
  • Recorded video streams

The system decodes video efficiently, runs the PPE detection model, tracks people across frames, applies PPE and zone rules, and generates an event when configurable persistence criteria are satisfied.

For example, a missing-helmet alert can require a person to remain inside a restricted area for several consecutive frames instead of triggering on a single frame affected by momentary occlusion.

PPE Detection Model Development and Optimization

In the process of model development, we will choose and analyze suitable object detection models, create industrially relevant datasets, annotate people and PPE categories, train or fine-tune models, and export them to ONNX and TensorRT to improve inference.

The selection of PPE categories depends on the purpose of the model. The validation of accuracy occurs via a reserved test dataset as well as footage collected from practice.

We perform a collection of explanations of classes, thresholds for acceptance, erroneous detections, and particular limitations of operation instead of presenting the laboratory accuracy as a guarantee for any camera or site.

Person Tracking, Zone Rules, and Event Detection

PPE detection becomes more useful when combined with person tracking and site-specific rules. We can configure virtual zones for restricted areas, entry points, production lines, loading areas, or other locations where PPE compliance is mandatory.

The system can apply rules such as:

  • Person enters a restricted zone without a helmet
  • Worker enters without a high-visibility vest
  • Required PPE is missing for a defined period
  • PPE compliance changes while a worker remains in a zone
  • Repeated violations occur within a configurable time window

Persistence rules help reduce unnecessary alerts caused by brief occlusions, camera noise, or temporary detection failures.

Web Dashboard and Safety-System Integration

For operations and safety teams, we develop a browser-based dashboard or integrate PPE events with an existing platform through REST APIs, MQTT, webhooks, or message queues.

Depending on requirements, the dashboard can display:

  • Camera status
  • Recent PPE events
  • Cropped evidence images
  • PPE category
  • Timestamp
  • Detection zone
  • Acknowledgement status
  • Escalation history

We can also implement role-based access control, audit logging, event-retention rules, CSV export, and notifications through email, Microsoft Teams, SMS gateways, or an existing site safety application.

On-Premise Privacy and Data Protection

Privacy is considered as part of the system architecture rather than as an afterthought. The Jetson-based solution can process camera streams fully on premises, reducing the need to transfer continuous video to a cloud platform.

Depending on organisational requirements, the system can:

  • Store event snapshots instead of continuous footage
  • Crop evidence to the relevant incident region
  • Apply face blurring where appropriate
  • Use short evidence-retention periods
  • Restrict remote access
  • Apply role-based permissions
  • Maintain audit logs for event access

For organisations using existing CCTV infrastructure, we work with security, legal, HR, and health-and-safety stakeholders to establish appropriate retention, access-control, signage, and review policies.

Edge AI can reduce unnecessary cloud video transfer, but it does not eliminate the need for a lawful and transparent workplace-monitoring policy.

Jetson Deployment, Monitoring, and Production Engineering

Our delivery covers the engineering work required to move a PPE detection solution from a development environment into an operational industrial deployment.

This can include:

  • Jetson device provisioning
  • Docker-based deployment where suitable
  • Secure system configuration
  • Application health monitoring
  • Watchdog and automatic restart behaviour
  • Model-version control
  • Camera and stream health checks
  • Acceptance testing
  • Production configuration
  • Deployment documentation
  • Engineering handover and support

We design the software stack with maintainability in mind so that models, application components, and configuration can be updated without unnecessarily disrupting the complete edge system.

Custom Industrial Hardware and Enclosure Integration

When the deployment requires hardware adaptation beyond a standard Jetson development kit, we can also support a custom carrier-board or enclosure strategy.

Depending on site requirements, this may include:

  • DIN-rail installation
  • Industrial power input
  • PoE networking
  • GPIO integration
  • UPS status monitoring
  • Temperature monitoring
  • Ruggedized field installation
  • Industrial enclosure design considerations

This enables the Jetson Orin Nano Super platform to be integrated more effectively into existing industrial infrastructure and edge-computing environments.

Why Choose an On-Premise Jetson PPE Detection System?

A Jetson-enabled solution installed on the premises offers industrial teams the ability for local processing of video information, allowing for a lesser dependency on Internet connection, easy integration into existing infrastructure, and better control over data processes.

This solution is especially useful when there are limitations in bandwidth, it is undesirable to use the cloud for video transfer, or when events with security implications need processing on-site with no delay.

Adequate Infosoft can deliver the full system: camera and network assessment, Jetson provisioning, edge-AI pipeline, model optimisation, dashboard and API development, privacy controls, field testing, documentation, and ongoing model improvement.

The result is a maintainable PPE detection platform designed around the realities of an industrial site, not a generic demo converted into a safety claim.

Editorial Resources

Ashok Patel
Ashok Patel
Senior Engineering Project Manager
AI/ML, DevOps, Data Science & Automation | IoT & C#/.NET | Azure & AWS Expert | Certified AI & Cloud Engineer | 1,500+ LinkedIn Followers
Facebook LinkedIn Stack Overflow