Skip to main content

Biometric matching

Biometric matching is at the heart of every Criminal Investigation deployment. This chapter focuses on the parts of the matching engine that are relevant for forensic work; the underlying generic concepts apply equally to other ABIS use cases.

Probe and reference​

Biometric matching always compares one probe with one or more references. In a criminal investigation:

  • the probe is normally the evidence collected at the crime scene (a latent print, a CCTV face, a video frame, a DNA profile),
  • the reference is normally a record stored in a criminal gallery (a known suspect, a wanted person), or another evidence in the criminal gallery (case-to-case search).

Matching score and threshold​

The matcher returns a matching score that quantifies how similar the two records are. The score is composed of:

  • partial scores – the score between two specific modality instances (e.g. left index against left index),
  • modality scores – aggregation of partial scores within one modality,
  • fused score – aggregation of modality scores across all modalities used in the request.

A pre-configured matching threshold decides whether a comparison is considered a match. The threshold is set by Innovatrics specialists during deployment based on the expected workload, the FAR/FRR requirements of the agency and the nature of the reference data.

Identification (1:N)​

Identification is the most common operation in criminal investigation: the probe is compared against a large gallery of references, and the matcher returns a ranked list of candidates above the threshold. The candidate count is bounded by a configurable parameter so that very crowded results are kept manageable.

Identification (1)

Galleries can be selected per request – an examiner can search only the wanted-persons gallery, only the criminal-applicants gallery, only the case gallery (i.e. evidence from previous cases), or any combination of them.

Verification (1:1, 1:few)​

Verification compares one probe against one (or a small enumerated set of) references. This is used for:

  • confirming a hit candidate – verifying the probe against a single reference returned by identification,
  • checking a witness statement – verifying a face captured at the location against a known suspect,
  • comparing two pieces of evidence – e.g. two latent prints from two different crime scenes.

Verification (1:1)

Stateless, semi-stateful and fully-stateful matching​

Identification and verification can be invoked in three modes:

  • Stateless – both the probe and the reference are sent in the request body. The records are not stored in ABIS. Useful for ad-hoc one-off comparisons that should not pollute the database.
  • Semi-stateful – the probe is sent in the request and the reference is referenced by ID (or vice versa). The stored side participates in the comparison, the other side is a transient input.
  • Fully-stateful – both the probe and the reference are stored in ABIS. The request only contains the IDs. The whole operation can be persisted as part of the audit log.

The fully-stateful mode is the recommended one for criminal investigation because it produces the richest audit trail.

Specialised forensic matching​

Beyond standard fingerprint, palm, face and iris matching the platform supports forensic-specific matching modes:

  • Latent-to-plain, latent-to-rolled and latent-to-latent matching with adjustable rotation (default 180°).

  • Eye-position normalisation for face comparison – the Comparison Tool aligns the two faces by eye position before the examiner reviews them.

  • Image normalisation for fingerprints / palms – the system rotates the two images so that they sit in parallel before the comparison.

  • Multimodal fusion combining face, fingerprint and iris – useful when the agency holds a multimodal record of a suspect and a multimodal probe is available.

    Multimodal matching

  • DNA kinship matching – returns persons whose DNA profile is related to the probe (parent, child, sibling, half-sibling).

  • SMT, body and face characteristic search – when biometric identification is not possible the system can search by NCIC-coded descriptors instead.

Search history​

Every fully-stateful matching request is persisted by the Investigation Service. The persisted record contains the request parameters (galleries, threshold, candidate count, rotation, etc.), the timestamp, the user, and the full candidate list with scores. Examiners can re-open the persisted search at any time.

API contract​

The REST contract for identification and verification is documented in the Identification and Verification API chapter.