Worker and machinery tracking
Multi-object tracking across overlapping camera coverage with consistent identity through occlusion.
AI intelligence system
SafePredict
From incident detection to incident prevention.
A vision system that models how people, vehicles and machinery move through a shared space, scoring emerging risk from trajectories and near misses instead of waiting for an incident to detect.
- Sector
- Industrial Safety
- Year
- 2024
- Duration
- 8 months
- Team
- 7 people
- Platforms
- Edge · Web console · Mobile alerts
- Status
- In production
3s
Prediction horizon
Trajectory projection window
Automatic
Near-miss capture
No reliance on voluntary reporting
On-edge
Inference location
No footage leaves the site
Anonymised
Worker tracking
Identity is not part of the model
01 — The problem
Safety programmes are built on incident reports, which means they learn from the events organisations most want to prevent. Near misses, which are far more frequent and equally instructive, were recorded only when somebody bothered to file one.
Existing camera systems detected a person in a restricted zone after entry. By then the useful window for intervention had closed. The client wanted to know about the convergence, not the collision.
02 — What we did
We model motion rather than presence. Workers, forklifts and machinery are tracked as trajectories, and risk is computed from projected paths — where these bodies will be in three seconds, given where they are heading now.
Near misses became first-class data. Every convergence that came within a defined margin is recorded automatically, giving the safety team a dense stream of leading indicators instead of a sparse trickle of lagging ones.
Waiting for an incident to learn from is the most expensive possible training set.
Interface
What it looks like in use
Two of the screens that carry the most weight in daily use, rebuilt here from the production design system.
Floor 2 · live riskELEVATED
Forklift 4 ↔ pedestrian · 2.8s to convergeALERT
Zone C · restricted · permit activeOK
Near misses today6
Anonymised trackingEdge inference
Risk rises before contactTrajectories are projected three seconds ahead, so convergence is flagged while there is still time to intervene.
Near-miss density · 30 daysLEADING INDICATOR
Aisle 7 crossing41
Dock 3 approach18
Press bay entry9
3sHORIZON
0FOOTAGE OFF SITE
Layout problems, not just behaviourNear-miss density mapped onto the floor plan separates badly arranged space from individual error.
Capabilities
What the system does
Grouped by the job each set of capabilities exists to do, rather than by which team built it.
Perception
01 · Seeing the floorAnonymous by design
No facial recognition and no worker identification. The model reasons about bodies and vehicles, not people.
Restricted-zone monitoring
Geofenced areas with rules that vary by time, machine state and permit status rather than being always-on.
Multi-camera fusion
Overlapping views fused into a single floor-plan coordinate space.
Risk modelling
02 · The predictive coreTrajectory prediction
Short-horizon path projection for every tracked body, accounting for speed, heading and typical routes.
Interaction analysis
Pairwise convergence risk between people, vehicles and moving machinery.
Near-miss detection
Convergences within a defined margin logged automatically with the clip and trajectory context.
Real-time risk scoring
A live floor-level score that rises before contact rather than reporting after it.
Response and analysis
03 · Closing the loopGraduated alerting
Escalation from local visual warning to supervisor notification based on score and persistence.
Hotspot analysis
Near-miss density mapped onto the floor plan, showing where layout rather than behaviour is the problem.
Shift and condition correlation
Risk patterns correlated with shift, congestion and production rate.
Intervention tracking
Whether a layout or process change actually reduced near-miss density, measured rather than assumed.
Architecture
How it is put together
Layer by layer, with the reason each one exists — because the reason is usually the interesting part.
01Edge inference
Detection and tracking run on site. Only anonymised trajectory data leaves the camera network.
- NVIDIA Jetson
- TensorRT
- Multi-object tracking
02Spatial fusion
Homography calibration mapping every camera view into a shared floor-plan coordinate system.
- Homography
- Calibration tooling
03Prediction
Short-horizon trajectory models with interaction-aware scoring across tracked pairs.
- Trajectory models
- Interaction scoring
04Alerting
Graduated notification with local signalling, supervisor push and escalation policy.
- MQTT
- Push notifications
05Analytics
Near-miss store with floor-plan heatmapping and intervention effectiveness tracking.
- PostgreSQL
- PostGIS
- React
Technology
Edge
- NVIDIA Jetson
- TensorRT
- GStreamer
- ONNX
AI & Vision
- PyTorch
- Multi-object tracking
- Trajectory models
Backend
- Python
- FastAPI
- PostgreSQL
- MQTT
Frontend
- React
- TypeScript
- Floor-plan rendering
Outcome
What changed
Measured against how the operation ran before, not against a benchmark chosen after the fact.
-
Leading indicators replaced lagging ones. The safety team works from near-miss density instead of incident counts.
-
Layout problems became visible. Hotspot mapping separated where the floor is badly arranged from where behaviour is at fault.
-
Footage never leaves the site. Inference runs at the edge and only anonymised trajectories are stored.
-
Interventions are measured. A change is judged by whether near-miss density actually fell.
How it ran
-
Months 1-2
Site study and consultation
Mapped movement patterns and agreed the privacy position with the workforce before any camera was configured.
-
Months 3-4
Perception and calibration
Tracking, multi-camera fusion and the floor-plan calibration tooling.
-
Months 5-6
Prediction and scoring
Trajectory models and interaction risk, tuned against historical near-miss footage.
-
Months 7-8
Alerting and analytics
Graduated alerting, hotspot analysis and intervention tracking.
Related work
From the same practice
Manufacturing & Production Quality
LineTrace
From defect detection to defect causality.
A multimodal system that correlates line video, machine telemetry, batch records and operator actions to explain why a defect happened, not merely that it did..
2025 · 10 months
Telecommunications & Network Operations
NetCause
Not alarm detection. Failure hypothesis reduction.
A correlation engine that turns an alarm storm into a short ranked list of probable root causes by reasoning over network topology and fault propagation rather than alarm counts..
2025 · 8 months
Oil & Gas · Asset Integrity
PipeIntegrity
From a single inspection to longitudinal asset integrity.
A vision system that detects and localises pipeline deterioration, then tracks each defect across inspection cycles so maintenance is prioritised by progression rather than by appearance..
2024 · 9 months
Something similar?
Running into the same problem SafePredict solved?
Tell us what you run. We will reply within two business days with what we would build for your situation — and, just as usefully, what we would leave out.