---
kind: concept
slug: biometric-identification
title: Biometric identification
summary: >-
  Identifying a person by comparing a probe with every enrolled template and returning the closest. One finger-knuckle edge-map method with three distances; author-reported: clipped Chamfer reaches 17.02% true positives on one 147-volunteer dataset. A gait dataset is listed but reports no identification result.
tags: [biometrics, identification, binary-images, similarity-measures, gait]
papers: ["2605.17673", "2502.17244"]
related: [chamfer-distance, biometric-template-protection, face-recognition, gait-recognition]
status: draft
updated_at: 2026-11-15
---

## Overview

Identification compares one probe with every enrolled template and returns the
identity of the closest one. Verification would instead check one claimed
identity. The only identification method here, a finger-knuckle print method,
takes the closest gallery finger as the match, scoring each comparison with a
distance between binary region-of-interest maps [arxiv:2605.17673 §III.D]. The
decision rule is the smallest distance, and the whole design lies in how the
template is built and which distance is used. A second card, a gait dataset,
is relevant as data for walking-based recognition but reports no identification
method or result [arxiv:2502.17244 § Background].

## Variants

### Edge-map templates with distance matching

Each finger image becomes two binary maps: a shadow map for grooves darker
than their neighbours and a light map for ridges brighter than their
neighbours [arxiv:2605.17673 §III.A] [eq. 1] [eq. 3]. A box-count filter
removes isolated pixels [arxiv:2605.17673 eq. 5], and a 220×110 region around
the knuckle is cropped from each map [arxiv:2605.17673 §III.C]. The thresholds and window
sizes are reported in the text of the noise and edge sections
[arxiv:2605.17673 §III.B]. Card: [A simple approach for biometrics: Finger-knuckle prints recognition based on a Sobel filter and similarity measures](/papers/2605.17673).

The three distances compared on these maps, using the paper's reported
results on one 147-volunteer finger-knuckle dataset [arxiv:2605.17673 tab. I]:

- **Mean absolute difference** [arxiv:2605.17673 eq. 6]. It is the time
  baseline in the paper (relative time 1) and has the lowest reported true-positive rate, 0.79% [arxiv:2605.17673 tab. I] [arxiv:2605.17673 tab. II].
- **Hausdorff distance** [arxiv:2605.17673 eq. 7]. Reported true-positive rate
  2.60%, about 224 times the baseline time [arxiv:2605.17673 tab. II]. The
  authors report that stray pixels can strongly inflate it [arxiv:2605.17673 §IV].
- **Chamfer distance clipped at τ** [arxiv:2605.17673 eq. 8]. Reported
  true-positive rate 17.02%, about 105 times the baseline time
  [arxiv:2605.17673 tab. I] [arxiv:2605.17673 tab. II]. Per finger, the
  true-positive rate ranges from 18.2% to 30.8% [arxiv:2605.17673 tab. III].
  See [chamfer-distance](/concepts/chamfer-distance) for background.

### Gait data for walking-based identification

The UNB StepUP-P150 dataset records 150 adults walking across a pressure
walkway at four speeds and in four footwear conditions, with raw 90-second
trials, two normalised footstep versions and per-step labels
[arxiv:2502.17244 § Methods]. The paper presents it as a resource for research
on recognising people by their walk, and notes that the scarcity of large public
pressure sets has limited comparison of deep-learning methods
[arxiv:2502.17244 § Background]. It trains and reports no identification model,
so it gives no accuracy or hit-rate figure for this concept. Its excluded
sessions are held back as independent test samples for future footstep biometric
competitions [arxiv:2502.17244 § Methods]. Card: [UNB StepUP-P150 plantar-pressure dataset](/papers/2502.17244).
For the gait side, see [gait-recognition](/concepts/gait-recognition).

Other biometric traits are covered by other pages:
[face-recognition](/concepts/face-recognition) (embeddings compared by
similarity) and [biometric-template-protection](/concepts/biometric-template-protection)
(measuring whether protected templates can be linked). None of these
has a card that describes the same method as the variant above.

## When to use which

| Situation | Use | Why |
|---|---|---|
| Lowest compute per comparison matters most, and a very low hit rate is acceptable | Mean absolute difference | Time baseline; true-positive rate 0.79% [arxiv:2605.17673 tab. I] [arxiv:2605.17673 tab. II] |
| Among the three measures, the best reported hit rate on binary edge maps matters, and about 100 times the baseline time is affordable | Chamfer distance clipped at τ | 17.02% true positives [arxiv:2605.17673 tab. I]; about 105 times baseline time [arxiv:2605.17673 tab. II] |
| Outlier sensitivity is a concern | Avoid Hausdorff as the only measure | Lowest hit rate after mean absolute difference and slowest; outlier-sensitive [arxiv:2605.17673 §IV] [arxiv:2605.17673 tab. II] |
| Need a public gait dataset with many footwear and speed conditions to build or test a walking-based identifier | UNB StepUP-P150 as data, not as a method | Covers 4 speeds and 4 footwear conditions for 150 adults [arxiv:2502.17244 § Methods]; no identification result is reported [arxiv:2502.17244 § Methods] |
| Data from many people across many sessions, or a need for a learned embedding | Not covered by this page's cards | The identification card uses fixed image processing and no training [arxiv:2605.17673 tab. I]; the gait card trains nothing |

Caveats for all rows. The results come from one dataset and one author-reported
setup [arxiv:2605.17673 §III] [arxiv:2605.17673 §IV]. The true-positive rate is
measured over all probe-gallery comparisons, and how that maps to a single
identification decision is not stated [arxiv:2605.17673 §IV]. Table IV sets this
paper's true-positive rate against accuracy figures from other papers, which use
different metrics and splits, so the rows are not directly comparable
[arxiv:2605.17673 tab. IV].

## Implementation sketch

Clipped Chamfer distance for one direction, the form the paper uses
[arxiv:2605.17673 eq. 8]. The paper does not say whether the two directions are
averaged, so this is one direction only.

```pseudocode
# Clipped Chamfer distance, one direction  [arxiv:2605.17673 §III.D] [eq. 8]
# Input:  marked-pixel point sets A, B (non-empty); clip tau (paper: 2)
# Output: mean clipped nearest-neighbour distance from A to B

def chamfer_clipped(A, B, tau):
    total = 0
    for p in A:
        # Nearest marked pixel in B, clipped at tau  [eq. 8]
        nearest = min(distance(p, q) for q in B)
        total = total + max(nearest, tau)
    return total / len(A)
```

Identification as a 1-to-N search over the gallery [arxiv:2605.17673 §III.D].

```pseudocode
# Identification: closest gallery entry  [arxiv:2605.17673 §III.D] [§IV]
# Input:  probe ROIs, gallery of (identity, ROIs), distance function
# Output: identity with the smallest distance

def identify(probe, gallery, measure):
    best_id = None
    best = infinity
    for identity, rois in gallery:
        # How the shadow and light scores are combined is not stated
        score = measure(probe, rois)
        if score < best:
            best = score
            best_id = identity
    return best_id
```

## Contradictions

None known. Only one card reports identification results, so no disagreement
between papers on identification performance can be checked yet. The gait
dataset reports no identification result, so its figures cannot be set against
the finger-knuckle numbers.

## Papers

- [A simple approach for biometrics: Finger-knuckle prints recognition based on a Sobel filter and similarity measures](/papers/2605.17673): edge-map templates for finger-knuckle prints, compared with mean absolute difference, Hausdorff or clipped Chamfer distance.
- [UNB StepUP-P150 plantar-pressure dataset](/papers/2502.17244): a public gait dataset for walking-based recognition research; reports no identification result.
