基于多分辨率卷积网络的交通标志检测与识别

发布时间：2018-04-01 08:18

本文选题：卷积神经网络　切入点：目标识别　出处：《大连理工大学》2015年硕士论文

【摘要】：在机动车通行道路上设置交通标志,对调节交通流量和提高道路的通行能力都起着重要的作用,它还可以预先把当前的道路状况向驾驶人员做出指示,让驾驶人员提前有所准备,从而降低交通事故发生的概率。但近年来,随着车辆数量的增加和道路状况的日益复杂,交通安全问题引发了人们的热议,智能交通系统也受到各方的关注,该系统通过综合运用先进的科学技术,建立一个可以主动控制车辆安全行驶的系统。交通标志识别作为该系统中辅助驾驶系统的一个重要功能,在整个系统中发挥了巨大的作用,各国政府和汽车制造商对这一问题进行了多年的研究,但由于真实交通环境的复杂性,还有许多问题尚待解决。为此,本论文对交通标志检测和识别这一问题作为研究主题,运用卷积神经网络的方法对其进行研究和分析。在交通标志检测问题上,为了提高从场景图像中检测交通标志的速度和准确度,本文提出一种非对称卷积神经网络的算法。该算法首先要从原始图像中提取ROI区域,而后进行分类,前期提取ROI区域的方法是用传统的颜色转换和形状匹配的方法进行,实验使用的原始图像来源于德国交通标志检测基准数据集。通过实验表明,该算法可以克服图像中交通标志尺寸过小、不良天气和标志褪色等不良影响,鲁棒性较好,证明了算法的有效性。在交通标志识别问题上,提出了一种多分辨率卷积神经网络算法,用于解决使用卷积神经网络进行交通标志识别时训练时间长的问题。该算法在基于传统卷积神经网络的基础上,使用两个不同的分支训练网络,提取特征并进行分类。将图像预处理为一个高分辨率图像数据集和一个低分辨率图像数据集,分别作为两个分支的输入,高分辨率分支可以更好的提取全局轮廓特征,低分辨率分支可以更好的提取局部细节特征,然后通过全连接层整合两个分支提取的特征,最后通过分类器进行分类识别。在GTSRB数据集上的实验表明,该算法在保证高识别精度的同时,大幅度降低了训练时间,兼顾了交通标志识别的准确性和实时性。
[Abstract]:The setting up of traffic signs on motor vehicle roads plays an important role in regulating traffic flow and improving the capacity of roads. It can also give directions to drivers in advance on the current road conditions.Make drivers prepare ahead of time, thus reducing the probability of traffic accidents.However, in recent years, with the increase of the number of vehicles and the increasing complexity of road conditions, traffic safety issues have aroused heated discussion, and the intelligent transportation system has also attracted the attention of all parties. The system through the comprehensive use of advanced science and technology,Establish a system that can actively control the safety of the vehicle.As an important function of the auxiliary driving system, traffic sign recognition has played an important role in the whole system. Governments and automobile manufacturers have studied this problem for many years.However, due to the complexity of the real traffic environment, there are still many problems to be solved.Therefore, this paper studies and analyzes traffic sign detection and recognition using convolution neural network.In order to improve the speed and accuracy of detecting traffic signs from scene images, an asymmetric convolution neural network algorithm is proposed in this paper.The algorithm firstly extracts ROI region from the original image and then classifies it. The previous extraction method of ROI region is based on the traditional color conversion and shape matching method.The original images used in the experiment come from the German traffic sign detection datum data set.Experiments show that the algorithm can overcome the bad effects such as too small size of traffic signs, bad weather and fading of signs in the image, and the robustness of the algorithm is good, and the effectiveness of the algorithm is proved.In traffic sign recognition, a multi-resolution convolution neural network algorithm is proposed to solve the problem of long training time in traffic sign recognition using convolution neural network.Based on the traditional convolution neural network, the algorithm uses two different branches to train the network to extract features and classify them.Preprocessing the image into a high-resolution image data set and a low-resolution image data set, respectively as the input of the two branches, the high-resolution branch can better extract the global contour features.Low resolution branching can better extract local detail features, and then integrate the features of the two branches through the full join layer, and finally classify and recognize by classifier.Experiments on the GTSRB dataset show that the algorithm not only ensures high recognition accuracy, but also reduces the training time and takes into account the accuracy and real time of traffic sign recognition.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;TP391.41

【共引文献】