Skip to main content

AI Detection System

Overview​

Built-in AI for Blue Iris v6 using ONNX runtime and GPU acceleration.

Runs object detection directly inside Blue Iris — no external AI server required.

The AI layer sits between motion detection and alerting:

Camera Feed -> Motion Trigger -> AI Confirmation -> Alert or Cancel

When motion fires, Blue Iris sends the frame to the AI engine.
If the AI confirms a recognized object above the confidence threshold,
the alert proceeds. If not, the alert is cancelled.

This is the primary mechanism for reducing false alerts on all cameras.

🎯 Objectives​

  • Confirm motion events using object classification before alerting
  • Reduce false positives caused by lighting changes, shadows, and foliage
  • Leverage GPU acceleration for low-latency inference
  • Support specialized detection models (animals, vehicles)

🧠 How Blue Iris Built-In AI Works​

Inference Flow​

  1. Camera detects motion
  2. Frame extracted from trigger
  3. Frame sent to ONNX engine
  4. Model classifies objects
  5. Results compared to camera filters
  6. Match above threshold -> alert confirmed
  7. No match -> alert cancelled

Key Concepts​

  • AI is per-camera
  • AI confirms, not replaces motion
  • Confidence threshold controls filtering
  • Object filters define triggers
  • "Cancel on no result" suppresses noise

What the AI Classifies​

  • Person
  • Vehicle
  • Animal (model-dependent)
  • Other YOLO-supported objects

⚡ GPU Acceleration with ONNX​

Runtime​

Blue Iris uses ONNX Runtime to load .onnx models directly.

Execution:

  • CUDA (preferred, NVIDIA)
  • DirectML (fallback)

Hardware​

ComponentValue
GPUNVIDIA RTX 4070 SUPER
RuntimeONNX Runtime
ExecutionCUDA

Why GPU Matters​

  • Faster inference per frame
  • Handles multiple cameras concurrently
  • Prevents missed detections
  • Keeps latency low

Verify GPU Is Active​

Blue Iris > Settings > AI

  • Provider shows ONNX + CUDA
  • Logs confirm initialization
  • CPU fallback = performance drop

📁 Model Configuration​

Models Folder​

C:\BlueIris\AI\models

Active Models​

ModelPurposeUse
yolov8s.onnxGeneral detectionPerson, vehicle
ipcam-animalAnimal detectionWildlife

yolov8s​

  • Fast, efficient
  • 80 COCO classes
  • Best default model

ipcam-animal​

  • Tuned for wildlife
  • Better outdoor accuracy
  • Assign per-camera

Adding Models​

  1. Place .onnx file in models folder
  2. Restart Blue Iris
  3. Select in AI settings

🎯 Confidence Tuning​

Current: 60%

RangeMeaning
90%+Very high certainty
70–89%Strong detection
60–69%Marginal
Below 60%Rejected

Tuning Strategy​

Too many false alerts

  • Raise to 65–75%

Missed detections

  • Lower to 50–55%

Per-Camera Guidance​

  • Close range -> 65–70%
  • Long range -> 50–55%
  • Animal zones -> 55–60%

🐾 Animal Detection​

Use Case​

Outdoor cameras:

  • Perimeter
  • Pool
  • Gate

Configuration​

  • Assign ipcam-animal
  • Keep yolov8s active
  • Enable animal class
  • Confidence: 55–60%

Behavior​

  • Animal detected -> alert
  • Human/vehicle still detected
  • Models run in parallel

⚙️ Optimization​

Max Connects​

Current: 8

ValueEffect
1–3Queue delays
8Balanced
16+GPU saturation risk

Frame Selection​

  • Trigger zone placement matters
  • Avoid edge detection
  • Use pre-trigger buffer

GPU Memory​

  • yolov8s = low footprint
  • Multiple large models increase VRAM usage
  • Monitor via Task Manager

Reduce AI Load​

  • Lower max connects
  • Use smaller model
  • Limit AI to key cameras
  • Improve motion filtering

Logging​

  • Use AI logs
  • Verify classifications
  • Watch for CPU fallback

📊 Reference​

Current Configuration​

SettingValue
AI ProviderONNX
GPURTX 4070 SUPER
ExecutionCUDA
Modelyolov8s
Animal Modelipcam-animal
Confidence60%
Max Connects8

Key Settings​

SettingLocation
AI providerSettings > AI
ModelSettings > AI
Max connectsSettings > AI
Camera AICamera > Trigger > AI
ConfidenceCamera > AI
LogsLog > AI

📊 Related Pages​

⚠️ Notes​

  • Requires Blue Iris v6
  • CUDA requires NVIDIA driver
  • CPU fallback reduces performance
  • ipcam-animal is supplemental
  • Confidence changes apply instantly
  • New models may require restart

✅ Result​

A GPU-accelerated AI system running fully inside Blue Iris v6.

yolov8s handles general detection.
ipcam-animal extends wildlife coverage.

60% confidence and 8 max connects provide stable multi-camera performance.