Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 44 of the 44 references it cites were found in a registry or an index; 2 of the 6 sources its claims cite were checked for support.

A High-Precision Vehicle Detection and Tracking Method Based on the Attention Mechanism

Open access

Abstract

Vehicle detection and tracking technology plays an important role in intelligent transportation management and control systems. This paper proposes a novel vehicle detection and tracking method for small target vehicles to achieve high detection and tracking accuracy based on the attention mechanism. We first develop a new vehicle detection model (YOLOv5-NAM) by adding the normalization-based attention module (NAM) to the classical YOLOv5s model. By exploiting the YOLOv5-NAM model as the vehicle detector, we then propose a real-time small target vehicle tracking method (JDE-YN), where the feature extraction process is embedded in the prediction head for joint training. Finally, we present extensive experimental results to verify our method on the UA-DETRAC dataset and to demonstrate that the method can effectively detect small target vehicles in real time. It is shown that compared with the original YOLOv5s model, the mAP value of the YOLOv5-NAM vehicle detection model is improved by 1.6%, while the MOTA value of the JDE-YN method improved by 0.9% compared with the original JDE method.

Overview

CoChat Summary

## TL;DR

This paper proposes a vehicle detection-and-tracking pipeline for small targets by adding a normalization-based attention module to YOLOv5s and combining it with a joint detection/tracking method. On UA-DETRAC, the authors report that detection mAP improves by 1.6%, tracking MOTA improves by 0.9%, and vehicle identity switches drop by 15%.

## Key findings

  • The proposed YOLOv5-NAM detector improves mAP over YOLOv5s by 1.6% overall on UA-DETRAC; the paper also reports 1.4% improvement from NAM alone and 0.2% more from SD-NMS.
  • For single-class detection on UA-DETRAC, the paper reports AP gains of 0.3% for cars, 2.7% for buses, 3.6% for other vehicles, while vans slightly decreased; with both NAM and SD-NMS, overall mAP improved by 1.6%.
  • The proposed JDE-YN tracking method improves MOTA by 0.9% compared with the original JDE method, and the paper says vehicle identity switching times decrease by 15%.
  • The paper reports the model after adding NAM increases network layers from 213 to 219 and model size from 13.7 M to 13.9 M, while FPS decreases by 0.4; replacing NMS with SD-NMS lowers FPS by 3 for YOLOv5s and 2.8 for YOLOv5-NAM.

## Methods

  • The paper uses a detection + tracking system model: YOLOv5-based vehicle detector plus a tracker built around Kalman filter state estimation and data association.
  • For detection, it modifies YOLOv5s by adding the NAM attention module to the Neck, uses CIoU_Loss instead of GIoU_Loss, and proposes Soft-DIoU NMS (SD-NMS).
  • For tracking, it proposes JDE-YN, embedding appearance feature extraction into the detection head and using a cascade matching algorithm based on direction correction.
  • Experiments are run on UA-DETRAC; the paper states it selects 16,542 train images / 102,562 labels and 11,235 test images / 86,559 labels for detection experiments, and for tracking uses 60 video sequences for training and 40 for testing.

## Limitations

  • The paper states the dataset is challenging because UA-DETRAC includes multiple weather conditions and vehicles of different scales, which keeps MOTA and MOTP low (between 0.3 and 0.4).
  • It explicitly notes a tradeoff: the proposed method improves accuracy but can reduce speed, e.g. FPS drops when adding NAM or SD-NMS. This is stated in the paper.
  • Evident from the text: the reported improvements are demonstrated mainly on UA-DETRAC, with a secondary COCO2017 retraining check; broader generalization beyond these datasets is not established in the provided text.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/s23020724

Cites

44 of 44 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 The detection-based vehicle tracking model.
Figure 1 The detection-based vehicle tracking model.CC BY · PubMed Central
Figure 2 The YOLOv5 network model.
Figure 2 The YOLOv5 network model.CC BY · PubMed Central
Figure 3 C3 structure.
Figure 3 C3 structure.CC BY · PubMed Central
Figure 4 SPPF structure.
Figure 4 SPPF structure.CC BY · PubMed Central
Figure 5 The FPN+PAN structure.
Figure 5 The FPN+PAN structure.CC BY · PubMed Central
Figure 6 Channel attention submodule of NAM.
Figure 6 Channel attention submodule of NAM.CC BY · PubMed Central
Figure 7 Spatial attention submodule of NAM.
Figure 7 Spatial attention submodule of NAM.CC BY · PubMed Central
Figure 8 The Neck of the YOLOv5-NAM network model.
Figure 8 The Neck of the YOLOv5-NAM network model.CC BY · PubMed Central
Figure 9 Vehicle overlapping.
Figure 9 Vehicle overlapping.CC BY · PubMed Central
Figure 10 Data association failure. (a) frame 115; (b) frame 187; (c) frame 132; (d) frame 188; (e) frame 137; (f) frame 190.
Figure 10 Data association failure. (a) frame 115; (b) frame 187; (c) frame 132; (d) frame 188; (e) frame 137; (f) frame 190.CC BY · PubMed Central
Figure 11 The cascade matching algorithm based on direction correction.
Figure 11 The cascade matching algorithm based on direction correction.CC BY · PubMed Central
Figure 12 The visual detection results of the YOLOv5s and the YOLOv5-NAM models.
Figure 12 The visual detection results of the YOLOv5s and the YOLOv5-NAM models.CC BY · PubMed Central
Figure 13 The vehicle target tracking results.
Figure 13 The vehicle target tracking results.CC BY · PubMed Central

Verification

Claims the paper makes

2 Valid · 2 Not valid · 4 Not checked · 45 Own results and interpretations

19 more claims

Create a free account to see all 53 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 22
  • The quote only partly states this claim. 1
  • Probably not a substantial claim. 8

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Cited by

See all 61

Metadata

Authors
Wang, J., Dong, Y., Zhao, S., & Zhang, Z.
Affiliations
Xidian University (CN)
Corresponding author
Zhao, Shuangrui
Venue
Sensors
Publisher
MDPI AG
Published
2023-01-08
Volume
23
Issue
2
Pages
724
ISSN
1424-8220 (ISSN-L)
OpenAlex
W4313894539
Cited by
60
References
44
Citations by year
2026: 6 · 2025: 16 · 2024: 26 · 2023: 12
FWCI
5.42
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Key R/D Program of China (2018YFE0207600, 61972308, 2019JC-17)
  • Natural Science Foundation of China (NSFC) (2018YFE0207600, 61972308, 2019JC-17)
  • Natural Science Basic Research Program of Shaanxi (2018YFE0207600, 61972308, 2019JC-17)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Surveillance and Tracking Methods
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Engineering › Automotive Engineering › Autonomous Vehicle Technology and Safety
Keywords
vehicle detection, vehicle tracking, small object detection, attention mechanism, YOLOv5, real-time tracking, intelligent transportation management, UA-DETRAC
MeSH terms
Intelligence, Technology, Transportation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.