Processing Pipeline

A source is registered (RTSP/ONVIF discovery, a file, or a looped clip). The reader opens the stream on its own thread and hands frames to a drop-oldest buffer so the analysis stage never blocks the camera.

Frames land in a fixed-size FrameBuffer (drop-oldest). A health monitor watches for freezes and reconnects. A rolling window of recent frames is retained for incident clips and appearance search.

Dark frames are auto-enhanced so faint subjects are still found; frames are resized to the working resolution. A background plate (EMA) accumulates the true static scene for later completion.

YOLO11 runs with per-class confidence floors and optional tiling. Disabled classes are dropped at the detector so nothing downstream pays for them. ByteTrack assigns stable ids inline.

Instance / foreground masks (YOLO11-seg) combine with a depth-derived foreground mask to decide which pixels are movers, so the background behind them can be reconstructed.

Depth Anything V2 produces relative inverse-depth, temporally median-fused across frames to suppress noise. The result is normalised to [0,1] and forms the geometry backbone.

A keypoint model runs on an interval over tracked people. Skeletons feed intent, posture cues and the gait descriptor used for cross-camera re-identification.

Depth is ground-plane fitted behind occluders; foreground masks and boxes mark movers; a completed background depth + texture is inpainted so the scene continues behind objects.

The depth grid is unprojected through a pinhole model into a coloured point cloud, with depth smoothing so surfaces stay continuous and discontinuities are ramped rather than torn.

The point grid is triangulated into a mesh; triangles straddling a depth discontinuity (object silhouettes) are culled, yielding clean edges for both the scene and the background layer.

The framed RGB, plus the inpainted background texture behind movers, is mapped onto the mesh, so occluded walls and floor render as real, textured surfaces.

Tracks gain velocity and predicted positions (foresight); contacts are placed on the depth-locked tactical radar; Re-ID links identities; the rule + threat engine raises correlated alerts.

Frames, detections, metrics and alerts stream over WebSocket to the UI. Scenes, incident clips, dossiers and event timelines can be exported. Everything persists locally in SQLite.