Department of Geomatics, National Cheng Kung University, Tainan City 701, Taiwan.
Sensors (Basel). 2022 Jun 24;22(13):4768. doi: 10.3390/s22134768.
In recent years, the development of self-driving cars and their inclusion in our daily life has rapidly transformed from an idea into a reality. One of the main issues that autonomous vehicles must face is the problem of traffic sign detection and recognition. Most works focusing on this problem utilize a two-phase approach. However, a fast-moving car has to quickly detect the sign as seen by humans and recognize the image it contains. In this paper, we chose to utilize two different solutions to solve tasks of detection and classification separately and compare the results of our method with a novel state-of-the-art detector, YOLOv5. Our approach utilizes the Mask R-CNN deep learning model in the first phase, which aims to detect traffic signs based on their shapes. The second phase uses the Xception model for the task of traffic sign classification. The dataset used in this work is a manually collected dataset of 11,074 Taiwanese traffic signs collected using mobile phone cameras and a GoPro camera mounted inside a car. It consists of 23 classes divided into 3 subclasses based on their shape. The conducted experiments utilized both versions of the dataset, class-based and shape-based. The experimental result shows that the precision, recall and mAP can be significantly improved for our proposed approach.
近年来，自动驾驶汽车的发展及其融入我们的日常生活已经迅速从一个想法变成了现实。自动驾驶汽车必须面对的主要问题之一是交通标志检测和识别问题。大多数关注这个问题的工作都采用了两阶段的方法。然而，快速行驶的汽车必须快速检测到人类所看到的标志，并识别出其中包含的图像。在本文中，我们选择使用两种不同的解决方案分别解决检测和分类任务，并将我们的方法与一种新颖的最先进的检测器 YOLOv5 的结果进行比较。我们的方法在第一阶段使用 Mask R-CNN 深度学习模型，旨在根据形状检测交通标志。第二阶段使用 Xception 模型完成交通标志分类任务。这项工作使用的数据集是一个手动收集的数据集，其中包含 11074 个台湾交通标志，这些标志是使用手机摄像头和安装在车内的 GoPro 相机收集的。它由 23 个类别组成，根据形状分为 3 个子类。实验分别使用了基于类和基于形状的两种数据集版本。实验结果表明，我们提出的方法可以显著提高精度、召回率和 mAP。
