An Attribute-Semantic-Enhanced, Zero-Parameter Feature Fusion Method for Vehicle Re-Identification

In suspect vehicle tracking and routine public-security surveillance, police forces often rely on multiple cameras to follow a target. Across cameras, however, viewpoint and illumination changes markedly degrade the discriminative power of purely visual vehicle re-identification (ReID) features, leading to recognition errors and reduced tracking efficiency. To address the problem of improving cross-camera matching accuracy at a low upgrade cost, this paper proposes an attribute-enhanced feature fusion method with zero training parameters: a 13-dimensional attribute probability vector, whose first eight entries are the predicted probabilities of eight vehicle-color categories (yellow, green, gray, red, blue, white, brown, and black) and whose last five entries are the probabilities of five vehicle-type categories (sedan, sport utility vehicle, bus, truck, and minivan), is scaled by a scene-adaptive weight λ, concatenated with the 128-dimensional ReID embedding, and L2-normalized (i.e., rescaled to unit Euclidean length), thereby completing the semantic injection in the cosine metric space. The fusion operator consists solely of a single scalar parameter and a normalization operation; it contains no trainable components, is fully decoupled from the ReID backbone, and supports plug-and-play use with any model. Experiments on the VeRi-776 (Vehicle Re-Identification, 776 identities) dataset show that the fusion method raises the overall mean average precision (mAP) from 38.64% to 42.41% (+3.77 percentage points, pp), improves cross-camera sub-scenario mAP by 4.77 pp, and yields a gain of +6.50 pp on the adjacent-camera sub-scenario. Zero-vector and random-attribute controls rule out the two confounds of dimensionality expansion and random effects, and the fusion gains on six heterogeneous ReID backbones are all positive (+3.16 pp to +5.84 pp). Mechanism analysis shows that the gain stems from the semantic regularization that attributes impose on cross-view retrieval ranking, and that a “substitution elasticity” exists between attributes and visual features: the more reliable the visual features, the smaller the marginal contribution of attributes. These results indicate that zero-parameter fusion achieves stable, attributable cross-camera matching gains at an extremely low upgrade cost, exhibits good model generality, and provides a technical reference for improving cross-camera vehicle tracking efficiency.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-18
DOI
https://doi.org/10.3390/app162211446
Primary Topic
Video Surveillance and Tracking Methods
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An Attribute-Semantic-Enhanced, Zero-Parameter Feature Fusion Method for Vehicle Re-Identification

Rongxi Zhang, Jie Shen, Jiyan Yu, Ruihan Ma
2 links
Applied Sciences
Video Surveillance and Tracking Methods
article

An Attribute-Semantic-Enhanced, Zero-Parameter Feature Fusion Method for Vehicle Re-Identification

Rongxi Zhang, Jie Shen, Jiyan Yu, Ruihan Ma
article en

Abstract

In suspect vehicle tracking and routine public-security surveillance, police forces often rely on multiple cameras to follow a target. Across cameras, however, viewpoint and illumination changes markedly degrade the discriminative power of purely visual vehicle re-identification (ReID) features, leading to recognition errors and reduced tracking efficiency. To address the problem of improving cross-camera matching accuracy at a low upgrade cost, this paper proposes an attribute-enhanced feature fusion method with zero training parameters: a 13-dimensional attribute probability vector, whose first eight entries are the predicted probabilities of eight vehicle-color categories (yellow, green, gray, red, blue, white, brown, and black) and whose last five entries are the probabilities of five vehicle-type categories (sedan, sport utility vehicle, bus, truck, and minivan), is scaled by a scene-adaptive weight λ, concatenated with the 128-dimensional ReID embedding, and L2-normalized (i.e., rescaled to unit Euclidean length), thereby completing the semantic injection in the cosine metric space. The fusion operator consists solely of a single scalar parameter and a normalization operation; it contains no trainable components, is fully decoupled from the ReID backbone, and supports plug-and-play use with any model. Experiments on the VeRi-776 (Vehicle Re-Identification, 776 identities) dataset show that the fusion method raises the overall mean average precision (mAP) from 38.64% to 42.41% (+3.77 percentage points, pp), improves cross-camera sub-scenario mAP by 4.77 pp, and yields a gain of +6.50 pp on the adjacent-camera sub-scenario. Zero-vector and random-attribute controls rule out the two confounds of dimensionality expansion and random effects, and the fusion gains on six heterogeneous ReID backbones are all positive (+3.16 pp to +5.84 pp). Mechanism analysis shows that the gain stems from the semantic regularization that attributes impose on cross-view retrieval ranking, and that a “substitution elasticity” exists between attributes and visual features: the more reliable the visual features, the smaller the marginal contribution of attributes. These results indicate that zero-parameter fusion achieves stable, attributable cross-camera matching gains at an extremely low upgrade cost, exhibits good model generality, and provides a technical reference for improving cross-camera vehicle tracking efficiency.

Applied SciencesVol. 16(22)
Nanjing University of Science and Technology (CN)
Video Surveillance and Tracking Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon