Skip to main content

INVESTIGATE · FACE & PLATE SEARCH

Find anyone in seconds. Find any vehicle the same way.

Drag in any photo — or pick a known subject — and Sentinel returns every ranked appearance across every camera in well under a second, even across millions of records. Vehicle and plate intelligence works the same way, with fuzzy matching that survives an imperfect read.

Face Search results grid: query photo top-left, ranked matches with similarity score, camera name, and timestamp; filter bar with date range and camera selectorClick to enlarge

Built for the first question after the incident: where have we seen this before?

Sub-second
Search at scale
every ranked appearance retrieved across millions of records
Any image
Upload search
a still, a phone shot, or a scanned photo — drag and drop
Adjustable
Similarity threshold
tune how strict a match must be, per case
Plate + vehicle
Fuzzy match
search by plate text that survives an imperfect read
Upload search

Drag in a photo. Retrieve every appearance.

Drag any still into Sentinel and it detects the face, generates its embedding, and returns every ranked appearance across your entire history in well under a second.

  • Any image works — drag and drop a still, a phone shot, or a scanned photo; no special format required
  • Faces found automatically — Sentinel detects the face and generates its signature before searching
  • A quality filter discards blur, extreme angles, and tiny crops so a poor input doesn't return noise
  • Ranked matches — results come back ordered by similarity, each with the camera, timestamp, and score
  • No name required — the photo is the query; identity, if any, comes later
Upload area with a dragged photo, detected face outlined, ranked results grid belowClick to enlarge

The photo is the query. Ranked matches in well under a second.

Plate & vehicle intelligence

Search by plate. Read the imperfect ones too.

Fuzzy plate matching surfaces near-misses alongside exact hits so a single misread character doesn’t lose the trail.

  • Text + fuzzy match — enter a full or partial plate; near-misses caused by an imperfect read surface alongside exact ones
  • Full sighting history — every reading across every camera, with camera, time, and the source frame
  • Cross-camera vehicle trace — follow a vehicle's path across the network by plate, the same way you follow a person
  • Region-correct rendering — the plate is displayed in its proper visual format for fast, unambiguous reading
  • Goes straight to a watchlist — a plate of interest can be added for ongoing alerting, with lawful basis recorded
Plate-search box with a query string, region-correct plate render, and a results panel listing sightings across cameras with timestampsClick to enlarge

A single misread character won't lose the trail.

From a search to ongoing monitoring

Found someone to keep watching? One step.

Sentinel won’t let an investigative lead quietly become a monitored subject off the books — a lawful basis is required before any watchlist entry is saved.

  • One step from any result — the face crop or plate and any known name carry over automatically
  • Lawful basis required at the point of action — a basis and its document reference and expiry must be recorded before saving
  • Suggested category — VIP, Person of Interest, Vehicle, and others, so the entry is filed correctly from the first second
  • Accountable from the start — the add is written to the tamper-evident audit
Add-to-watchlist panel opened from a search result: face crop carried over, name pre-filled, category suggested, lawful-basis dropdown and document-reference field requiredClick to enlarge

A lawful basis is required — there is no "save it later."

The full search toolkit

Every way to find, filter, and act.

Known-subject search

When the subject is already in your records, pick them from the list — no photo needed. The search runs against their established signature and returns their complete appearance history across every camera and time range you allow. One step from the profile.

Similarity ranking & threshold

Every result carries a similarity score; the list is ordered strongest-first. A single threshold control lets you decide how close a match must be to make the list at all — tighten it for confidence, loosen it when you're casting a wider net. The operator always confirms before a result becomes an identity.

Time & camera filters

Filter results down to the time range and the cameras that fit your case before you start reading: a single afternoon, a specific entrance, a group of cameras covering one site. Stack time and camera filters with the similarity threshold to go from thousands of candidates to the handful that matter.

Search at scale

Sentinel stores each face as a compact mathematical signature and indexes them for approximate-nearest-neighbour retrieval, so a query stays sub-second whether your history holds hundreds of thousands of records or millions. No external search cluster, no GPU required for the search itself.

Plate & vehicle search

Type a full or partial plate; fuzzy matching surfaces near-misses caused by an imperfect read, so a single wrong character doesn't lose the trail. Every reading across every camera, with the full sighting history attached. Follow a vehicle's path the same way you follow a person.

Add to watchlist in one step

From any result — a face match or a plate — add the subject to a watchlist for ongoing alerting. The crop or plate carries straight across, the name pre-fills if known. A lawful basis must be recorded before the entry is saved; there is no "save it later." Every add is audited.

How it works

From a photo to a pattern.

1

Search

Drag in a photo, pick a known subject, or type a plate. Sentinel detects the face or reads the plate, then searches your entire history in well under a second.

2

Refine

Order is strongest-first by similarity. Tighten the threshold for confidence or loosen it for reach, then narrow by time range and camera until you're left with the appearances that matter.

3

Act

Open any result to play the clip, follow the subject's path across cameras, or open their Pattern-of-Life — and add a face or plate of interest straight to a watchlist, with lawful basis, for ongoing alerting.

4

Account for it

Every search and every watchlist add is written to a tamper-evident audit, so the work is accountable from the first query, not reconstructed afterward.

Under the hood

Specifications

Face detectionSCRFD (ONNX), with YuNet as a lighter alternate
Face embeddingArcFace, 512-dimensional (ONNX)
Vector storePostgreSQL with pgvector
Index typeIVFFlat on cosine distance
Search methodApproximate nearest neighbour (kNN)
Search latencySub-second across millions of embeddings (index capability, measured on representative data)
Similarity scoringCosine similarity, ranked, with an adjustable match threshold
FiltersDate/time range · camera · camera group
Face quality filterDiscards blur, extreme angle, and faces under ~40 px before embedding
Plate detection & OCRPaddleOCR detection + recognition (ONNX); fuzzy text match; region-correct UI render
Inference runtimeONNX Runtime 1.22.0
Search infrastructureNo external search engine; no GPU required for search
Add-to-watchlistLawful basis required at point of action; audited
AccountabilityEvery search and watchlist add written to the tamper-evident audit

Specifications describe shipped platform capabilities; we'll confirm the configuration that fits your deployment during your demo.

Find a person of interest on live data.

Request demo access and we’ll send you a private, pre-loaded environment. Drag in a photo, watch every ranked appearance come back in well under a second, follow the subject across cameras, and add a result to a watchlist — the full search workflow, end to end.