Software Configuration

1. Detection (What to watch)

Configure the sources the detector should monitor.

• source RTSP/HTTP stream URL or local video path. Use a list [ ] for multiple sources.
• interval Seconds to wait between processed frames (default 0). Useful to reduce load on slow machines.
• frame_retention How many recent frames to keep in memory per source (default 30) for context.

2. YOLO (Object Detection)

The fast first-pass AI that scans every frame. Without YOLO, all frames simply pass through.

• model URL or local path to a YOLO model file (.pt or .onnx).
• confidence Confidence required (0-1). Can also be a per-class threshold, e.g., undefined.
• time_max Maximum duration in seconds to group frames into one event (default 60).
• timeout Seconds of no detections before the event is considered over (default 5).
• cooldown Seconds to wait before starting a new event. Prevents repeat alerts. Can be per class.
• include_trailing_time Seconds of extra footage to include after the last detection (default 1).
• frames_min How many frames in a row must match before counting (default 6, or 3 depending on GPU).
• imgsz Image size fed into the model. Higher is more accurate but slower (default 640).
• strategy Which frames to evaluate: "LATEST" (most recent only) or "ALL".

3. VLM (AI Double-Check - Optional)

A Vision Language Model checks the footage after a YOLO detection. Greatly reduces false alarms.

• prompt The question to ask, e.g., 'Is there a person in this video?'.
• model The AI model (e.g., 'gemini/gemini-2.0-flash'). Supports LiteLLM providers.
• key API key for the chosen model provider.
• url Custom API endpoint, useful for local models.
• strategy "VIDEO" sends a clip to the AI. "IMAGE" sends a single frame (cheaper but less accurate).

4. Exporters (Alerts & Storage)

Configure where confirmed detections should be sent.

Disk (Opslag)

• directory Folder path under 'detections/' to save files into. Otherwise uses the class name.
• strategy "BEST" saves only the highest-confidence frame. "ALL" saves every frame.
• confidence Minimum confidence required to save.
• export_rejected Whether to also save detections rejected by the VLM (default true).

Telegram

• token Your Telegram bot token.
• chat The Telegram chat or user ID.
• confidence Minimum confidence required to send.
• alert_every Only send an alert every Nth detection (1 = every time, 5 = every 5th).
• include_plot Include the full frame with a detection box drawn on it (default false).
• include_crop Include a cropped image of just the detected object (default false).
• include_video Include an MP4 clip of the detection sequence (default true).
• video_width Width of the video clip in pixels (default 1280).
• video_crf Video quality (0-51). Lower = better quality, default 28.
• export_rejected Whether to also send detections rejected by the VLM (default false).

Webhook

• url The URL to POST to when a detection occurs.
• token Authorization token sent in the request headers.
• confidence Minimum confidence required to trigger.
• data_type How image/video data is encoded in the payload: "binary" or "base64".
• data_max Maximum payload size in bytes. Images are compressed if they exceed this.
• include_plot Include the full frame with detection overlay (default false).
• include_crop Include a cropped image of the detected area (default true).
• include_video Include an MP4 clip of the detection sequence (default false).
• video_width Width of the video clip in pixels (default 1280).
• video_crf Video quality (0-51). Lower = better quality, default 28.
• export_rejected Whether to also POST detections rejected by the VLM (default false).

5. Health (Ping Monitor)

Sends a simple periodic HTTP request to verify the process is still alive.

• url The URL to ping.
• method HTTP method: GET, POST, PUT, PATCH, DELETE, or HEAD (default GET).
• interval Seconds between pings (default 60).
• timeout Request timeout in seconds (default 5).
• headers Optional HTTP headers map.
• body Optional request body sent as raw text.

6. ONNX Runtime

Controls how ONNX Runtime configures before loading a YOLO model.

• provider Optional provider name to force (e.g., "CUDAExecutionProvider").
• winml For Windows ML build only. Tries to register ML execution providers (default true).
• opset ONNX opset used when exporting. Lower values can improve compatibility (default 20).