Edgegenix Edge AI runs sub-40ms vision models directly on the asset — smoke, wildfire, object detection, and thermal fire mapping — fully offline. Detects. Decides. Reports when connected.
From raw sensor input to verified finding — every step runs locally on the Edgegenix AI Box.
RGB, thermal, and depth frames from camera or drone feed at up to 60fps
Hardware-accelerated normalise, resize, and channel fusion before inference
YOLO-class models run on-device NPU — sub-40ms end-to-end latency
Confidence scoring, class labelling, bounding box generation, and deduplication
Verified findings timestamped and forwarded to Cloud AI when network available
Edge AI is purpose-built for environments where cloud dependency is not an option.
Detects smoke plumes and fire ignition from fixed towers, vehicles, and drones. On-device fusion with weather and fuel load data. First alerts before a 000 call is made.
SMOKE_V3 · 82% mAPClassifies vehicles, people, wildlife, and equipment with >95% accuracy. Computes collision and intrusion paths entirely on-device for safety and access control.
4 CLASSES · >95% mAPReads and classifies road signs, speed limits, and regulatory markings in real time. Deployed in autonomous vehicle trials and remote mine site haul roads.
SIGN_V2 · 94% mAPDual RGB+thermal input maps active fire zones, hotspot progression, and safe evacuation corridors — displayed in real time on truck-mounted screens without network.
THERMAL · 38ms LATENCYSMOKE_V3 runs a YOLO-class detector fused with wind speed, relative humidity, and fuel dryness index. Every detection includes spread-risk scoring, not just a bounding box.
Camera or drone frame at up to 4K
Wind, humidity, fuel dryness co-read from onboard sensors
On-device YOLO detection model — 38ms
Spread probability and confidence score computed locally
Finding dispatched to Cloud AI and emergency contacts
The on-device risk engine fuses real-time camera output with localised weather data pulled from onboard sensors — eliminating false positives from dust, steam, and haze without a network call. When spread risk exceeds threshold, the finding is written to local storage with full timestamp, GPS, and environmental context, then forwarded to Cloud AI when connectivity returns.
Deployed across tower networks in NSW and WA, the system has triggered verified alerts up to 47 minutes before the first 000 call in field trials.
Primary ignition detected. Wind ENE 28km/h. Fuel dryness index 4.1. Spread probability HIGH.
OBJ_V4 classifies four primary object classes with >95% mean average precision at 60fps. Intrusion path and collision risk are computed on-device — no cloud lookup required.
Wide-angle camera at full resolution
Crop, normalise, channel stack via NPU
Multi-class YOLO on NPU — 22ms
Intrusion and collision vector on-device
Triggered alert + timestamped finding stored
SIGN_V2 reads and classifies regulatory, warning, and advisory road signage in real time — deployed in autonomous vehicle trials and remote mine site haul roads at >100km/h approach speed.
Forward-facing wide-angle RGB input
Sign classifier on NPU — 28ms
Speed value and text extraction
Speed limit fed to vehicle control bus
Reads numeric speed values from standard and non-standard sign formats including zone markers
Detects sharp curves, intersection ahead, animal crossing, and school zone markers
Stop, give way, no entry, one-way — all classified with occlusion tolerance up to 40%
Custom sign classes for haul road intersections, blast zones, and exclusion areas
Classifies lane types, chevrons, and painted road instructions alongside signage
Operates under headlight illumination with IR-enhanced input for 24-hour coverage
Dual RGB+thermal cameras on the truck feed THERM_V2, which maps active fire perimeters, hotspot progression, and dynamically computes safe evacuation corridors — all without connectivity.
Two active hotspots detected. Perimeter expanding NE at approx 12m/min. Escape route generated: NW exit via Track C.
Dual RGB+FLIR thermal array feeds the Edge AI Box mounted in cab — no external power, runs on vehicle 12V.
Live operational picture. Aggregates findings from all assets, drones, and sensors into one verified intelligence feed for command.
Explore Cloud AI →On-asset vision AI. Runs fully offline on the Edgegenix AI Box — smoke, objects, thermal, signs — sub-40ms, no cloud required.
You are hereConnected intelligence layer. Integrates field findings with control room systems, emergency dispatch, and enterprise data platforms.
Explore ECIA →Request a live demonstration on your hardware or environment — wildfire, thermal, object detection, or custom use case.
Request a live demo