Turning Existing CCTV Feeds into Verified Early-Wildfire Events
The first visible indication of a wildfire is often a small, faint smoke plume occupying only a tiny part of a wide landscape image.
Detecting that plume is difficult. Acting on it is even more difficult.
Clouds, haze, glare, compression artefacts and brief camera movement can resemble smoke. A system that generates too many uncertain alerts will quickly lose the trust of the people expected to use it.
Our goal is therefore not simply to display a YOLO bounding box. We want to demonstrate how edge AI can convert an ambiguous visual signal into a persistent, explainable and operationally useful wildfire event.
What we will build
fAIrefighter Metis will be a CCTV-based early-wildfire monitoring prototype running on the Dell Pro Slim Plus XE5 and its 16 GB Axelera Metis PCIe accelerator.
The completed prototype will process one live CCTV camera through an RTSP stream. Its software architecture will be camera-source independent so that additional feeds can be configured in a future deployment, but the challenge deliverable will remain deliberately focused on a reliable, fully demonstrated single-camera system.
The system will contain four principal stages:
- High-resolution wildfire-smoke detection on Metis.
- Standard temporal tracking and event confirmation.
- Environmental context from an off-the-shelf weather station.
- An operator dashboard presenting live and historical events.
High-resolution smoke detection on Metis
Smoke detection will use YOLO11s Sensitive Detector, developed specifically for early wildfire smoke detection.
The model operates at a fixed input resolution of 1024 × 1024. Preserving this resolution is important because an early smoke plume may appear as a very small, low-contrast feature in a wide CCTV view.
The original model weights will be adapted and quantized for execution on the Metis platform through the Voyager SDK. We will validate the Metis version against the original model using the same representative footage, rather than assuming that successful compilation automatically preserves detection behaviour.
The comparison will examine detection agreement, small-smoke sensitivity, bounding-box behaviour, end-to-end latency, system throughput and the effect of host-side video processing.
From detections to verified events
The project will not add a second LSTM or image-classification network after YOLO.
Instead, it will use a conventional and transparent tracking-by-detection workflow. Every potential smoke region will be observed over successive frames. A configurable temporal policy will evaluate its persistence, confidence history and spatial consistency before changing its state from an unconfirmed candidate to a confirmed smoke event.
Brief or inconsistent predictions will expire without becoming alerts. Confirmed events will also use a separate clearing policy to prevent unstable on/off behaviour.
This approach is computationally practical, understandable to an operator and achievable within the challenge period.
Weather as operational context
A low-cost commercial weather station will be installed or represented as a remote data source alongside the camera system.
Depending on the selected device interface, the application will ingest measurements such as:
- Temperature.
- Relative humidity.
- Wind speed.
- Wind direction.
- Atmospheric pressure.
- Rainfall.
Weather information will not be used to claim that smoke has been detected. Visual inference and temporal confirmation will make that decision.
Instead, weather will provide context for understanding the possible urgency and development of a confirmed event. For example, a persistent visual smoke event combined with low humidity and strong wind may warrant greater operator attention than the same detection under less hazardous conditions.
A simulated weather-data source will also be provided so that the complete project can be reproduced without requiring the exact commercial station.
The dashboard
The primary user-facing deliverable will be a local web dashboard.
It will display:
- The live CCTV feed.
- Smoke bounding boxes and current confidence.
- Candidate, confirmed and cleared event states.
- How long a candidate has persisted.
- Camera identity and location metadata.
- Current weather conditions.
- Metis inference and pipeline performance.
- A timeline of previous events.
- Event snapshots or short evidence clips.
- Camera, weather-station and pipeline health.
Each confirmed event will create an evidence record showing what was detected, when it first appeared, when it was confirmed and what environmental conditions were present.
The demonstration will focus on the dashboard rather than LTE, SMS or emergency-service integration. Those are possible future outputs, but they are outside the challenge prototype’s scope.
Why Axelera Metis
Continuous analysis of high-resolution CCTV video is a natural edge-AI workload.
Running the detector locally avoids the need to upload a permanent video stream to a cloud inference service. The Metis accelerator will execute the neural-network workload, while the host will manage video decoding, temporal event logic, weather ingestion, evidence storage and the dashboard.
The project will also explore the practical limits of a fixed 1024 × 1024 specialist detector on the challenge hardware. This includes identifying whether inference, video decoding, preprocessing or postprocessing is the true end-to-end bottleneck.
The result will be a realistic foundation for future installations in which one Metis-equipped station processes feeds from multiple existing CCTV cameras.
How Voyager Wingman will be used
Voyager Wingman will be a central development tool and part of the documented experiment.
We will use it to:
- Develop the custom YOLO11 model-conversion process.
- Preserve and verify the model’s 1024 × 1024 preprocessing.
- Generate and refine the RTSP Voyager pipeline.
- Configure YOLO decoding and non-maximum suppression.
- Add tracking and expose detection metadata to the application.
- Investigate unsupported operations or CPU fallback.
- Profile complete pipeline performance.
- Generate reproducible setup and deployment instructions.
- Help diagnose differences between the original and Metis model outputs.
We will retain the initial prompt, important follow-up prompts, generated files, failures, corrections and validation results. This will show not only the final pipeline, but how Wingman helped turn the idea into working Metis software.
Validation
The prototype will be evaluated using both positive smoke footage and difficult negative footage.
Testing will include examples involving:
- Small or distant smoke.
- Clouds and haze.
- Fog.
- Strong sunlight and glare.
- Video compression artefacts.
- Temporary camera movement.
- Short-lived false detections.
We will compare operation with and without temporal confirmation and conduct an extended-duration stability test.
The final report will separate model inference performance from complete system performance and will document any accuracy change introduced by quantization.
Why we can complete it
We are not beginning with an untested problem area.
The project builds on previous work involving wildfire-detection research, camera-based inference and environmental sensing. For this challenge, we have intentionally reduced the scope: one specialist model, one real CCTV stream, one standard temporal-verification layer, one weather interface and one complete dashboard.
The main technical work is clearly bounded:
- Adapt and validate the model on Metis.
- Build the RTSP inference pipeline.
- Implement event persistence and weather ingestion.
- Deliver the dashboard, testing and documentation.
Four-week delivery plan
Week 1 — Baseline and pipeline
Establish the original model baseline, install and verify the Axelera environment, create the RTSP pipeline and document the initial Wingman interaction.
Week 2 — Metis model adaptation
Compile and quantize the YOLO11s detector, reproduce its preprocessing and postprocessing, compare output with the original model and resolve compatibility issues.
Week 3 — Verified events and dashboard
Add tracking, temporal event states, weather-station ingestion, evidence capture and the first complete dashboard workflow.
Week 4 — Validation and publication
Run difficult-negative and extended-duration tests, measure model and end-to-end performance, record the demonstration and publish the reproducible project materials and Wingman journey.
Final deliverables
The final submission will include:
- A working RTSP CCTV demonstration running on Metis.
- The adapted YOLO11s deployment pipeline.
- Temporal candidate and event-confirmation logic.
- Weather-station and simulated-weather integration.
- A live operational dashboard.
- Accuracy and performance comparisons.
- Difficult-negative and stability-test results.
- A demonstration video.
- Source code, configurations and setup documentation.
- The complete publishable Voyager Wingman prompt journey.
Intended impact
The long-term opportunity is to add intelligent wildfire monitoring to CCTV infrastructure that may already exist at forests, utilities, municipalities, industrial sites and protected areas.
By analysing high-resolution video locally, the system can convert continuous camera footage into compact, evidence-backed events. That allows operators to review what matters without requiring all video to be continuously transferred to a cloud inference system.
The challenge prototype will demonstrate the most important first step: turning one CCTV feed into a reliable, explainable and reproducible wildfire-monitoring workflow on Axelera Metis.
