基于显著性与卷积神经网络的交通标志检测与识别研究

发布时间：2018-06-18 07:49

本文选题：交通标志检测 + 交通标志识别　；参考：《长安大学》2017年硕士论文

【摘要】：随着科技的迅猛发展,汽车成为日常生活中人们出行必不可少的交通工具,由于机动车数量的持续增长,一系列交通问题随之而来,比如交通安全、交通拥堵、交通污染等。在这复杂的交通问题背景下,高级驾驶辅助系统(ADAS)应运而生。交通标志检测与识别作为ADAS的一个基础分支,也是提高交通安全和效率的重要手段。所以本文对交通标志检测与识别方法进行了相关研究。论文首先分析了显著性检测原理,结合交通标志的相关特征,提出一种基于显著特征的交通标志检测方法,并对该方法进行实验验证;其次,通过对卷积神经网络的分析和研究,提出一种基于改进的AlexNet卷积神经网络的交通标志识别方法。本文主要研究内容及成果包括以下几个方面:(1)提出一种基于显著性的交通标志检测方法。首先通过对交通标志的颜色特征、边界特征和位置信息这三个视觉特征进行显著性建模;其次建立显著特征融合规则,并结合代价函数的最小化,融合输入图像的特征,得到最终最优显著图;最后,对最优显著图进行二值化处理,提取并标记二值图中的连通区域,将其映射到原始RGB图中,使用滑动窗口法提取感兴趣区域,实现交通标志的检测。实验结果证明本算法适用于复杂环境下的交通标志的检测,并通过与4种常用的显著检测算法对比分析,证明本文提出的检测算法相比较于其它算法具有较高的显著性检测性能。(2)提出一种基于改进的AlexNet卷积神经网络的交通标志识别方法。该方法首先分析了AlexNet网络模型的结构,并在此网络的基础上对网络的结构和参数进行调整优化,得到新的AlexNet网络模型;然后使用新的网络模型对交通标志进行识别。利用改进后的AlexNet卷积神经网络进行交通标志的识别主要包括两部分内容,一是AlexNet网络模型的训练;二是利用训练好的AlexNet模型实现对输入的分类。(3)提出一种训练数据集扩充方法。本文选用德国交通标志识别数据集(GTSRB)对提出的AlexNet模型进行训练和测试,由于GTSRB训练样本的不平衡,本文提出两种样本扩充方法对数据集进行改善。实验采用扩充后的数据集和原始数据集对提出的Alex Net模型进行训练和测试,结果表明,使用扩充训练样本集训练的AlexNet分类模型对交通标志识别,测试集中大部分类别的交通标志能够达到95%以上的识别准确率,高于原始训练集的93%,通过实验,对比分析本文提出的Alex Net网络、LeNet卷积神经网络和经典的“Hog+SVM”分类器,证明本文提出的识别方法,无论是识别精准率还是时间复杂度方面,均优于另外两种方法。
[Abstract]:With the rapid development of science and technology, automobile becomes an indispensable vehicle in daily life. Because of the continuous growth of the number of vehicles, a series of traffic problems, such as traffic safety, traffic congestion, traffic pollution and so on. In the context of this complex traffic problem, Advanced driving Assistance system (ADASS) emerged as the times require. As a basic branch of ADAS, traffic sign detection and recognition is also an important means to improve traffic safety and efficiency. Therefore, this paper carries on the correlation research to the traffic sign detection and the recognition method. In this paper, the principle of significance detection is analyzed, and a new method of traffic sign detection based on salient features is proposed, which is verified by experiments. Based on the analysis and research of convolution neural network, a traffic sign recognition method based on improved AlexNet convolution neural network is proposed. The main contents and achievements of this paper include the following aspects: 1) A signal-based traffic sign detection method is proposed. Firstly, the visual features of traffic signs, such as color features, boundary features and location information, are modeled significantly. Secondly, the fusion rules of salient features are established, and the features of input images are fused with the minimization of the cost function. The final optimal salience map is obtained. Finally, the connected region in the binary map is extracted and marked, and mapped to the original RGB map, and the region of interest is extracted by sliding window method. The detection of traffic signs is realized. The experimental results show that the proposed algorithm is suitable for the detection of traffic signs in complex environments. It is proved that the proposed detection algorithm has higher significant detection performance than other algorithms.) A traffic sign recognition method based on improved AlexNet convolution neural network is proposed. Firstly, the structure of AlexNet network model is analyzed, and the structure and parameters of the network are adjusted and optimized on the basis of this network, and a new AlexNet network model is obtained, and then the new network model is used to identify traffic signs. The traffic sign recognition based on the improved AlexNet convolution neural network includes two parts: one is the training of the AlexNet network model, the other is to use the trained AlexNet model to realize the classification of the input. In this paper, the German Traffic sign recognition dataset (GTSRB) is used to train and test the proposed AlexNet model. Due to the imbalance of GTSRB training samples, two methods of sample expansion are proposed to improve the data set. The extended data set and the original data set are used to train and test the proposed Alex net model. The results show that the AlexNet classification model trained by the extended training sample set is used to recognize traffic signs. Most types of traffic signs in the test set can achieve more than 95% recognition accuracy, which is higher than 93% of the original training set. Through experiments, the paper compares and analyzes the Alex net network LeNet convolution neural network and the classical "Hog SVM" classifier. It is proved that the method proposed in this paper is superior to the other two methods in terms of accuracy rate and time complexity.
【学位授予单位】：长安大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】