自然场景车标数据集的构建及其应用

车标作为车辆身份的关键特征之一,在车辆的监控与辨识中发挥着重要作用。由于自然场景复杂多变,对其中的车标进行准确识别仍具有很大的挑战性。目前公开数据库很少且存在诸多局限,导致研究缺乏可信度和实用性。本文建立了一个面向自然场景的全新数据集,包含多种采集环境下的10 324幅、67类车辆图像。基于此数据集开展应用研究,提出一个目标检测与深度学习相结合的车标识别方法,包括车标区域定位和车标种类预测两大步骤。实验表明,该方法对复杂背景有较强的适应性,在涉及30种车标的分类任务中达到89.0%的总体识别率。

Construction of vehicle logo dataset in natural scenes and its application

As one of the key characteristics of vehicle identity, vehicle logo plays an important role in vehicle monitoring and identification. Due to the complexity of the natural scene, it is still a great challenge to identifying the vehicle logo accurately. At present, there are few open databases and there are many limitations, which lead to the lack of credibility and practicability. In this paper, a new dataset for natural scenes which contains 10 324 images with 67 types of vehicle logos in various acquisition environments was established. Based on this dataset, a vehicle logo recognition method based on target detection and deep learning was proposed. The method includes two major steps:regional positioning of vehicle logo and prediction of vehicle logo type. Experiments show that the proposed method has strong adaptability to complex background, and the overall recognition rate reaches 89.0% in the classification task involving 30 kinds of vehicle logos.