Detection Configuration

AI detection is configured at two levels: server-level settings that control the detection engine, and camera-level toggles that control which cameras participate in detection.

Server-Level Configuration

Server-level detection settings are managed in Settings > System. These settings apply to the detection engine running on each recording server.

Enabling Detection

Detection must be explicitly enabled for each recording server that has a GPU. Navigate to Settings > System and locate the recording server in the server list. Toggle the AI Detection switch to enable detection on that server.

Caution

Detection works best on recording servers with a supported hardware accelerator (NVIDIA GPU, Hailo, or Coral). Enabling detection on a server without acceleration will fall back to CPU inference, which is too slow for production use with many cameras.

Model Selection

Meridian VMS supports multiple detection model sizes, each offering a different balance between speed and accuracy:

Model	Size	Speed	Accuracy	Recommended For
Nano	~6 MB	Fastest	Lower	Large camera counts, limited GPU
Small	~22 MB	Fast	Good	General surveillance
Medium	~52 MB	Moderate	Better	Balanced deployments
Large	~88 MB	Slower	Best	Critical detection, fewer cameras
VisDrone	Varies	Moderate	Specialised	Aerial/drone footage, small objects
Combined	Varies	Slower	Broad	Mixed environments

Select the model that best fits your deployment:

• For deployments with many cameras (100+), use Nano or Small to maintain detection throughput.
• For deployments where detection accuracy is critical and camera count is moderate, use Medium or Large.
• VisDrone is optimised for detecting small objects in wide-angle or aerial views.
• Combined merges multiple model strengths for environments with diverse detection needs.

Model changes

Changing the model requires the detection engine to restart and load the new model weights. There will be a brief interruption in detection (typically 10—30 seconds) while the new model loads.

Inference FPS

The Inference FPS setting controls how frequently the detection engine captures and analyses frames. The configurable range is 0.5 to 10 FPS.

FPS	Behaviour
0.5	One frame every 2 seconds — lowest GPU load
2.0	Default — good balance for most deployments
5.0	Higher detection responsiveness, increased GPU load
10.0	Maximum — suitable for critical areas with powerful GPU

Tip

The effective per-camera FPS also depends on the total camera count and pool size. With 160 cameras at 2 FPS cycle rate, each camera is sampled at approximately 0.5 FPS regardless of the FPS setting.

Confidence Threshold

The Confidence Threshold sets the minimum detection confidence required for an event to be generated. The configurable range is 10% to 90%.

• Lower values (10—30%) produce more detections but increase false positives.
• Higher values (60—90%) produce fewer detections but with higher certainty.
• The default is 50%, which provides a reasonable balance for most environments.

Note

Heatmap data is collected at a lower confidence threshold (30%) regardless of this setting. This ensures heatmaps capture broad activity patterns even when the event threshold is set higher.

Class Selection

Select which object classes should trigger detection events. Only selected classes will generate events; unselected classes are ignored during event processing (though they may still contribute to heatmap data).

Available classes:

• Person
• Car
• Truck
• Bus
• Motorcycle
• Bicycle
• Dog
• Cat
• Bird

The default selection includes person, car, truck, bus, motorcycle, and bicycle — the classes most relevant to surveillance.

Camera-Level Configuration

Per-camera detection settings are managed in Settings > AI Detection. This page displays a tree view of cameras, filtered to show only cameras assigned to recording servers with detection enabled (GPU-equipped servers).

Detection Cameras Tree

The detection camera tree mirrors the location hierarchy but only displays:

• Locations that contain cameras on GPU-enabled servers
• Cameras whose recording server has detection enabled

This filtered view makes it straightforward to manage detection across large deployments without scrolling through cameras on servers that do not support detection.

Per-Camera Detection Toggle

Each camera in the tree has a detection toggle. When disabled:

• The camera is excluded from the detection engine’s capture rotation
• No detection events are generated for that camera
• Recording and live streaming continue unaffected
• Pool capacity is freed for remaining detection-enabled cameras

Optimising detection throughput

Disable detection on cameras where it is not needed (e.g. static indoor views, parking structure overviews with no security value). Fewer cameras in the detection pool means each remaining camera is sampled more frequently, improving detection responsiveness.

When to Disable Camera Detection

Common scenarios for disabling per-camera detection:

• Indoor corridors or lobbies where detection is not operationally useful
• Cameras with persistent false positives that cannot be resolved with zone or confidence adjustments
• Cameras pointed at foliage or water that cause excessive motion-triggered false detections
• Temporary disablement during maintenance or camera repositioning

Configuration Summary

Setting	Location	Scope	Default
Detection enabled	Settings > System	Per server	Disabled
Model	Settings > System	Per server	Small
Inference FPS	Settings > System	Per server	2.0
Confidence threshold	Settings > System	Per server	50%
Class selection	Settings > System	Per server	person, car, truck, bus, motorcycle, bicycle
Camera detection toggle	Settings > AI Detection	Per camera	Enabled