| 摘 要: 一直以来，如何准确便捷地监测能见度都是减少交通事故工作关注的重点所在，而传统的使用能见度仪检测能见度存在造价高、范围小等缺陷。随着深度学习研究的不断发展，使用深度学习估算能见度也变成现实。本文建立了一种VGGnet 16卷积神经网络模型，经过Dropout(丢弃层)和数据增强优化后，使用监控图像及其对应的能见度值对其进行训练。结果表明，优化后的模型能有效提升训练精度，在小数据集上也能实现非常好的能见度估算效果。 |
| 关键词: 能见度；深度学习；卷积神经网络；丢弃层；数据增强 |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 上海市自然基金项目(1818ZR1416500)；曙光计划(19SG51). |
|
| Visibility Estimation based on Convolutional Neural Network |
|
CAO Shuangliang, YANG Yali, CHEN Hao, YANG Shuwei
|
(School of Mechanical and Automotive Engineering, Shanghai University of Engineering Science, Shanghai 201620, China)
caoshuangliang2020@163.com; carolyn71@163.com; pschenhao@163.com; 18016334293@163.com
|
| Abstract: Accurate and convenient visibility monitoring has always been the focus of reducing traffic accidents. Traditional visibility meter has the effect of high cost and small range. With the continuous development of deep learning research, deep learning is used to estimate visibility in reality. This paper proposes to establish a VGGnet 16 convolutional neural network model. After Dropout (discard layer) and data enhancement optimization, monitoring image and its corresponding visibility value are used to train it. The results show that the optimized model can effectively improve training accuracy, and it can also achieve very good visibility estimation effect on small data sets. |
| Keywords: visibility; deep learning; convolutional neural network; discarding layer; data enhancement |
