Deep Learning for Robust Vehicle Identification

Ramajo Ballester, Álvaro; González Cepeda, Jacobo; Armingol Moreno, José María · 2022 · Crossref

DOI: 10.1007/978-3-031-21065-5_29

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of robust vehicle identification in video surveillance systems, specifically targeting scenarios where license plate characters are partially occluded or distorted. While Automatic License Plate Recognition (ALPR) is widely used, existing solutions often fail under such adverse conditions. To overcome these limitations, the authors propose a "two-factor authentication" approach that combines traditional license plate reading with visual vehicle re-identification. This dual-method strategy aims to enhance system versatility and reliability by leveraging both textual data from plates and visual features of the vehicle body. The methodology employs a cascaded deep learning architecture. The first stage utilizes YOLOv5 for object detection and Optical Character Recognition (OCR). The authors developed and trained these models using a custom Spanish ALPR Dataset (SAD), optimizing hyperparameters such as batch size, image input dimensions, and optimizers to improve detection and character recognition accuracy. The second stage implements FastReID for vehicle re-identification, which extracts visual embeddings from vehicle images. This component was evaluated using both public datasets (Stanford Cars, VeRi) and proprietary datasets: the Highway Gantry Dataset (HGD) and the Operational Urban Dataset (OUD). The system integrates these neural networks to provide identification through either plate reading or visual matching, depending on data availability and quality. The results demonstrate that the combined system effectively addresses the shortcomings of single-method identification. The YOLOv5 modules achieved high performance in vehicle and license plate detection, as well as character recognition, with specific optimizations yielding improved metrics compared to baseline configurations. The FastReID module successfully re-identified vehicles based on visual characteristics, even when plate information was unavailable or unreliable. Comparative tests against standard models showed that the proposed architecture provided greater robustness in real-world conditions, such as varying lighting and occlusion, by relying on the complementary strength of the two identification factors. The significance of this work lies in its contribution to intelligent video surveillance for security and crime prevention. By integrating ALPR with deep learning-based vehicle re-identification, the system offers a more resilient solution for tracking vehicles in complex environments. This approach reduces dependency on clear license plate visibility, thereby improving the overall effectiveness of surveillance systems. The findings suggest that multi-modal identification strategies, combining textual and visual data, are superior to single-source methods for robust vehicle tracking, offering practical implications for real-time security applications and automated monitoring systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms