AI analytics
Object detection, counting, heat maps, faces, plates, and fire/smoke — on the box.
CoreSight runs AI analytics on the box itself — no cloud service, no footage leaving your site. Analytics run as separate worker processes, so heavy inference never destabilizes recording or live view.
Enabling AI
AI is opt-in at install time with the --ai flag:
curl -fsSL https://coresightvms.com/install.sh | sudo bash -s -- --prod --aiThis stages the inference runtime and an object-detection model. Offline install bundles built with AI assets enable it automatically. To add AI to a box installed without it, re-run the installer with --ai (re-running is non-destructive — see Updating).
Without AI provisioned, CoreSight degrades gracefully: nothing crashes and every non-AI feature works normally. The analytics pages simply have no detections to show. Face recognition, ALPR, and fire/smoke each need their additional models staged on the box — contact support if a feature you expect shows no results.
Inference runs on CPU by default; boxes with a supported GPU (CUDA, TensorRT, OpenVINO) use it automatically for higher camera counts, with CPU fallback.
Object detection
The foundation: cameras are analyzed for objects — people, vehicles, and other common classes. Detections are stored with the camera, time, class, and confidence, and feed everything below. Browse and filter them on the Detections page.
Smart search with clip-jump
The Detections page is searchable: filter by camera, object class, and time range ("person, loading dock, 22:00–06:00"). Every result deep-links to the exact recording — one click and you are watching the moment in Playback.
Tracking and counting
Detected objects are tracked across frames, which enables counting:
- Line-cross counts — draw a line (a doorway, a gate); CoreSight counts objects crossing it, per direction.
- Zone counts — draw a zone; CoreSight counts entries and occupancy.
Counts are reviewed on the Counts page, per camera and time range — footfall, vehicle throughput, occupancy over the day.
Heat map
The Heatmap page renders where activity concentrates in each camera's view over a chosen period — the at-a-glance answer to "where do people actually walk?". Useful for retail layout, queue analysis, and spotting unusual activity patterns.
Face recognition and identities
With the face models provisioned, CoreSight detects faces and matches them against an identity list you manage. A match becomes an event carrying the identity — usable in alarms and in the access-control allowlist. Matching happens entirely on the box.
License plate recognition (ALPR)
Vehicle plates are detected and read automatically, including Arabic plates (normalized to Latin characters so one search finds both). Plate reads are searchable events and can drive access control (gate opening for allowlisted plates).
Fire and smoke detection
Cameras can be analyzed for visible fire and smoke, producing events that feed the alarm workflow — an early-warning layer alongside conventional detectors, not a replacement for them.
Analytics and alarms
Any analytics event — a detection, a face match, a plate read, fire/smoke — can raise an alarm with escalation and notifications, exactly like motion. See Alarms and events.