.png){kind=icon}
.png){kind=icon}
返回
Data Augmentation and Intelligent Recognition in Pavement Texture Using a Deep Learning
.png){kind=icon}
PRE
AI
DOI：10.1109/TITS.2022.3140586.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
As one of the most important properties of road, the adaption to roads with different macro-textures may significantly affect the autonomous driving technologies since road texture directly affects the skidding resistance and tire noise. Therefore, it is of great significance to detect and analyze the road macro texture with respect to different pavement types and service conditions. Generally, transportation engineers may face problems such as small dataset size, unbalanced dataset, etc. To solve these problems, this study aims to recognize the pavement texture using the deep learning approaches. The pavement texture data was first visualized using image processing methods, and then augmented using the traditional methods as well as a deep learning approach, i.e. Generative Adversarial Network (GAN) model. The Random Forest (RF) algorithm and the DenseNet network were both employed, where the overall classification accuracy of the original dataset was 50% and 59%, respectively, and the accuracy of the data augmented by the traditional methods was 58% and 70%, respectively. Test results show that, after 250,000 generations of training, GAN model was able to generate new pavement texture images with high quality, and the classification accuracy on the test dataset using DenseNet improved to 82%. It was discovered that the deep learning methods had a better performance for pavement texture recognition than manual classification and traditional machine learning methods. Furthermore, it was also found that adding noise in the original datasets as an augmentation method had a negative impact on the classification accuracy.
Keyword:
Deep learning
Training
Roads
Feature extraction
Random forests
Generative adversarial networks
Data models
Data augmentation
pavement texture
recognition and classification
deep learning
autonomous driving
期刊
IF:
8.4
论文数:
9.5K
被引数:
6.3W