Running YOLOv8 on NVIDIA Jetson: Edge AI for Industrial Operations

Why Edge AI Matters for Industrial Operations

The typical cloud vision pipeline looks like this: capture a frame, compress it, upload it over a cellular or satellite link, wait for a cloud GPU to run inference, receive the result, and act on it. The round trip takes somewhere between 200 milliseconds and several seconds depending on connectivity. For a consumer photo-tagging application, that is perfectly acceptable.

For industrial operations, it is not. When a person walks into a restricted zone on a mine site, the system needs to detect them and trigger an alarm in under 100 milliseconds — not two seconds later. When a vehicle approaches a gate, the license plate needs to be read before the vehicle reaches the barrier. When wildlife enters a solar farm perimeter, the deterrent system needs to activate immediately, not after the animal has already reached the panels.

Beyond latency, there are three other compelling reasons to run inference at the edge:

• Bandwidth — A single 1080p camera at 30 FPS generates roughly 150 Mbps of raw video. Even with H.265 compression, streaming a dozen cameras to the cloud over a satellite link is prohibitively expensive. Edge inference processes frames locally and sends only metadata (bounding boxes, classifications, confidence scores) to the cloud — reducing bandwidth by 99%.
• Privacy — Many industrial facilities operate under strict data sovereignty requirements. Video of personnel, vehicles, and facility interiors cannot leave the premises. Edge inference means the raw video never touches the internet.
• Cost — Cloud vision APIs charge per frame or per API call. At 30 FPS across 10 cameras, you are looking at 26 million frames per day. Even at the lowest cloud pricing tiers, the cost is staggering. Edge hardware is a one-time capital expense.

NVIDIA Jetson Compute Tiers

The NVIDIA Jetson platform provides a range of compute options that scale from basic detection to multi-model inference pipelines. Powoflow's Mesa N1 edge device is built around the Jetson ecosystem, offering three hardware tiers matched to operational requirements:

• Jetson Orin Nano (26 TOPS) — The entry point for edge AI. Sufficient for running a single YOLOv8 model at 30 FPS on one or two camera streams. Ideal for gate monitoring, perimeter detection, or single-zone PPE compliance.
• Jetson Orin NX (67 TOPS) — The mid-range option. Supports running multiple models simultaneously — for example, person detection on four cameras plus license plate recognition on a gate camera. Handles multi-stream inference without frame drops.
• Jetson AGX Orin (100+ TOPS) — The high-end tier for complex deployments. Supports six or more camera streams with multiple concurrent models, face recognition pipelines, and real-time analytics overlays. Used in high-security facilities and large industrial complexes.

The TOPS (Tera Operations Per Second) metric directly translates to how many frames per second you can process at a given model complexity. A YOLOv8 nano model requires roughly 0.5 TOPS per camera stream at 30 FPS. A full YOLOv8 large model with face recognition adds significantly more compute demand. Matching the hardware tier to the workload avoids both over-provisioning and performance bottlenecks.

Five Detection Models for Industrial Use

Powoflow ships five pre-trained detection models optimized for industrial environments. Each model is trained on industrial-context datasets — not generic internet images — and fine-tuned for the lighting conditions, camera angles, and object types found on mine sites, energy facilities, and construction zones.

Person and PPE Detection

Detects people in the frame and classifies their PPE compliance — hard hat, high-visibility vest, safety glasses, gloves. Non-compliant detections trigger configurable alarms. This model is the most widely deployed across Powoflow customers and runs reliably on all three Jetson tiers.

Vehicle Detection and Classification

Identifies vehicles by type (light vehicle, heavy vehicle, truck, forklift) and tracks movement direction. Used for traffic management on mine haul roads, gate access logging, and exclusion zone enforcement.

Face Recognition

A two-stage pipeline using RetinaFace for face detection and ArcFace for embedding extraction. Detected faces are matched against an enrolled personnel database to provide identity verification at access points. The entire pipeline runs on-device — face embeddings are stored locally, and no facial images are transmitted to the cloud.

License Plate OCR

A two-stage pipeline using LPDNet for plate detection and LPRNet for character recognition. Supports multi-format plates across different regional standards. Detected plate numbers are matched against authorized vehicle lists for automated gate access. Recognition accuracy exceeds 97% under normal lighting conditions.

Wildlife Monitoring

Detects and classifies animals in and around facility perimeters. Used primarily on solar farms and wind installations for environmental compliance monitoring, deterrent activation, and wildlife corridor analysis. The model distinguishes between birds, medium mammals, and large mammals to trigger appropriate responses.

TensorRT Optimization

Raw PyTorch models are too slow for real-time edge inference. Every model deployed to the Mesa N1 is compiled through NVIDIA TensorRT, which applies layer fusion, kernel auto-tuning, precision calibration, and memory optimization specific to the target Jetson hardware.

The performance difference is dramatic. A YOLOv8 medium model running in native PyTorch on a Jetson Orin NX achieves roughly 8 FPS. The same model compiled with TensorRT in FP16 precision runs at 30+ FPS — a nearly 4x improvement with negligible accuracy loss. For models where FP16 is insufficient, TensorRT supports INT8 quantization with calibration datasets to maintain accuracy while further improving throughput.

TensorRT compilation is performed once per model-hardware combination. Compiled engines are cached on the device and loaded at startup, eliminating the compilation overhead during normal operation.

The Edge-to-Cloud Pipeline

Detection at the edge is only useful if it connects to your operational workflows. The Mesa N1 does not just detect objects — it publishes structured events that flow through the entire Powoflow platform.

The pipeline works as follows:

• The Jetson device runs inference on camera frames and generates detection events with bounding boxes, classifications, confidence scores, and timestamps.
• Detection events are published onto the platform event bus over the device's authenticated MQTT connection.
• Routing rules deliver detections to the alarm engine, which evaluates them against configured alarm rules (e.g., “person detected in Zone 3 after hours”).
• Triggered alarms generate push notifications to relevant personnel and appear in the Powoflow alarm queue.
• If configured, the alarm can automatically create a work order for investigation, complete with the detection snapshot and metadata.

The entire chain — from frame capture to push notification — completes in under 3 seconds on a typical LTE connection. The detection itself happens in under 50 milliseconds at the edge.

Deploying a model to the fleet

Deploying and updating AI models on hundreds of edge devices is an operational challenge. Powoflow uses a managed edge component runtime for over-the-air model deployment, enabling new models and model updates to be pushed to devices in the field without physical access.

When a new model version is released — whether it is a retrained detection model with improved accuracy or an entirely new model type — the deployment process is straightforward. The compiled TensorRT engine is packaged as a versioned component, published to the component registry, and deployed to target devices or device groups. The runtime handles the download, verification, and rollback-on-failure logic.

This capability is particularly important for wildlife models, which are often fine-tuned for specific regional species. A solar farm in the Australian outback needs a different wildlife model than one in the American Southwest. OTA deployment makes regional model customization practical without sending technicians to each site.

Cost Comparison: Edge vs Cloud

The economics of edge AI versus cloud vision APIs are straightforward once you run the numbers.

Consider a modest deployment: 10 cameras running at 15 FPS, analyzed 24 hours a day. That is 12.96 million frames per day, or roughly 389 million frames per month.

• Cloud vision API — At $1 to $3 per 1,000 frames, the monthly cost ranges from $389,000 to $1.17 million. Even with aggressive batching and sampling (analyzing every 10th frame), the cost is $39,000 to $117,000 per month. This is a perpetual operating expense that scales linearly with camera count.
• Edge hardware — A Jetson Orin NX module costs approximately $500 to $700. The complete Mesa N1 unit, with enclosure, storage, and connectivity, ranges from $500 to $2,000 depending on configuration. This is a one-time capital expense. The device processes every frame, at full frame rate, with no per-frame cost.

The payback period for edge hardware versus cloud APIs is typically measured in days, not months. For any deployment with more than a handful of cameras running continuous analysis, edge inference is the only economically viable option.