在智能交通系统中，车辆作为最普及的交通工具，常被不法分子利用使其成为一种安全隐患，因此实现监控设备下的车辆身份识别一直是一个研究热点。车标是车辆的特殊身份，其中包含着车辆品牌制造商的基本信息，相比车牌、车型和车色，车标具有相对独立和可靠的特性。车辆标志识别能够快速、精准地缩小车辆查询范围，为案件侦破、交通自动化管理等有效降低搜索成本，因此车辆标志识别在车辆身份识别中显得尤其重要。本文对近十年内的主流车标识别方法进行了系统概述，为车标识别领域内的后续研究者提供参考。首先，简要阐述了在智能交通系统中车标识别技术的研究背景和重要性。其次，根据车标识别过程中是否依赖手工提取特征，将目前国际主流的车标识别方法归纳为传统的车标识别方法和基于深度学习的车标识别方法，并分别总结了这两类方法的优劣。随后，分类、梳理和评价了这两类方法中现有的各种算法。再次，针对车标数据集稀少导致难以权衡各类算法性能、阻滞车标识别研究进展的问题，详细介绍了四种公开车标数据集：XMU（Xiamen University Vehicle Logo Dataset）、CarL-CNN 数据集、HFUT-VL（Vehicle Logo Dataset from Hefei University of Technology）和VLD-45（Vehicle Logo Dataset-45），并给出下载地址，可供研究者进行实验和测试。此外，描述了四种常用的评价指标，并在公开数据集上基于这些评价指标对车标识别方法开展实验，并对实验结果进行比较和分析。最后，综述现有车标识别技术中存在的一些问题与挑战，对未来车标识别的研究方向做出预测和展望。

Abstract: In intelligent transportation systems (ITS), vehicles, as the most popular means of transportation, are often used by lawless elements to make them a security risk, so the realization of vehicle identification under the monitoring equipment has been a research hotspot. The vehicle logo is the special identity of the vehicle, which contains the basic information of the vehicle brand manufacturer, compared with the license plate, model, and color of the vehicle, the vehicle logo has a relatively independent and reliable characteristic. Vehicle logo recognition quickly and accurately narrows down the scope of vehicle search, making it particularly important in vehicle identification. This paper presents a systematic overview of the mainstream vehicle logo recognition methods from the last decade to provide a reference for subsequent researchers in the field. First, it is explained that it is continuously under construction and development. Vehicle identification strongly supports the development and maturity of ITS. Vehicle identity consists of four parts: vehicle logos, license plates, vehicle models, and vehicle colors. To reduce algorithmic costs and improve the accuracy of vehicle identity recognition, vehicle logo recognition implementation is the most suitable for current needs. Secondly, the current international mainstream methods for vehicle logo recognition can be categorized into classical and deep learning-based approaches, depending on whether they rely on manual feature extraction. This section summarizes the advantages, disadvantages, and main ideas of both types of methods. Classical methods for recognizing vehicle logos can design proprietary solutions for specific vehicle logo recognition problems. The methods have the benefit of minimal dependence on the number of training samples and low hardware requirements, but they require manual feature extraction and cannot learn vehicle logo features independently for automatic recognition. The classical method for recognizing vehicle logos involves the following steps: inputting the image, performing preprocessing operations, extracting features, recognizing vehicle logos, and outputting the final result with accuracy. The vehicle logo recognition based on deep learning methods circumvents the laborious manual feature extraction process and performs better when sufficient samples are available. However, it incurs higher computational costs and demands more advanced hardware. The main approach of this method entails the creation of a vehicle logo recognition module and a model training module using deep learning techniques. In the logo recognition module, inputting the logo image is required followed by performing preprocessing operations on the image. Logo recognition is then accomplished through the application of deep learning methods, and the final performance is the accurate output of the recognition results. In the model training module, it is essential to prepare a substantial dataset, apply preprocessing operations, connect the neural network structure for independent learning and feature extraction from vehicle logo images, and utilize the classification network to accomplish the recognition and classification of vehicle logos. These two methods are further subdivided into contemporary international mainstream techniques. Among them, classical vehicle logo recognition methods can be categorized into four types: those based on scale-invariant feature transform (SIFT) feature extraction, histogram of oriented gradient (HOG) feature extraction, in-variant moments, and other classical recognition methods. Additionally, vehicle logo recognition based on deep learning methods can be divided into three types: those based on you only look once (YOLO) series of algorithms, deep residual network (ResNet) algorithms, and other algorithms based on convolutional neural networks (CNNs). This paper systematically sorts out the characteristics, advantages, and disadvantages of various algorithms as well as the datasets used in these methods. Again, addressing the problem that the scarcity of datasets for vehicle logos makes it challenging to evaluate the effectiveness of different algorithms and hinders research on recognizing vehicle logos, we provide a detailed explanation of four publicly available vehicle logo datasets. The XMU (Xiamen University Vehicle Logo Dataset), CarL-CNN dataset, HFUT-VL (Vehicle Logo Dataset from Hefei University of Technology), and VLD-45 (Vehicle Logo Dataset-45) are available for researchers to conduct experiments and tests via the provided download address. In addition, four commonly used evaluation metrics are described, and experiments on vehicle logo recognition methods based on these evaluation metrics are carried out on a publicly available dataset, and the results are compared and analyzed. Finally, although conventional methods of vehicle logo recognition perform well in small sample environments and many solutions have been proposed for certain complex environments, they still have limitations when faced with complex and variable traffic situations. Although the use of a deep learning-based vehicle logo recognition method leads to improved recognition and robustness of the model after model training, this improvement comes at the cost of training on a large-scale vehicle logo dataset and the need for constant hardware updates. By synthesizing the problems and challenges faced by classical vehicle logo recognition methods in ITS and vehicle logo recognition based on deep learning methods, this paper presents the following predictions and future development directions: 1) Developing new algorithms for low-cost, highly robust, and efficient vehicle logo recognition to meet practical applications. Vehicle logo recognition is a common image classification problem in the complex traffic environment. This task inevitably faces severe challenges from factors such as lighting effects, inclination changes, occlusion, wear and tear, and extreme weather. The development of new algorithms that balance recognition accuracy and speed while reducing costs and complexity, thereby expanding deployment scenarios of the model, remains a worthy research direction for continuous exploration. 2) Dynamic video research broadens the scope of vehicle logo recognition applications. The current reliance on static images for vehicle logo recognition presents challenges in data acquisition and expansion, consuming time and resources and limiting scalability and efficiency. Dealing with multi-vehicle scenarios and continuous dynamic scenes adds complexity. Dynamic video-based methods capitalize on the advantages of easily collectible video data, capturing vehicle logos from diverse angles and environments. Consequently, video-based vehicle logo recognition opens avenues for future research with both new opportunities and challenges. 3) Integrating the Transformer visual model enhances network structure to boost performance. In recent years, Transformer neural networks have gained attention for their exceptional representational ability and efficient processing of global information, showing promise in recognition tasks. In contrast to CNNs, Transformer visual models excel in image comprehension, global attention, and mitigating feature loss. Thus, incorporating Transformer visual models into vehicle logo recognition research is of substantial value. 4) Combining artificial intelligence (AI) large models enhances cross-modal open-domain vehicle logo recognition by integrating multimodal data for improved model robustness and accuracy. This approach integrates vehicle logo features with associated textual data, such as manufacturer and model number, in a unified model to address limited multimodal information challenges. AI large models effectively tackle data scarcity in recognizing cross-modal open-domain decals, extracting richer patterns from limited data to enhance recognition of unknown categories. Despite their powerful capabilities, deploying these models for vehicle logo recognition in open-domain scenarios poses cost challenges, making it a complex and cutting-edge task.