基于特征融合与深度卷积神经网络的交通标识识别

发布时间：2018-03-05 14:39

本文选题：交通标识识别　切入点：特征融合　出处：《广东工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：汽车在人们生活中扮演着越来越重要的角色,安全畅通的驾驶环境是交通系统的理想状态。交通标识识别是智能交通系统的重要组成部分,它主要包括交通标识的目标定位和目标识别两部分。以交通标识为研究对象,提出了基于特征融合和深度卷积神经网络的交通标识识别方法。首先介绍了国内外交通标识识别的研究现状,对比了过去研究中目标定位与目标识别方法的优劣,提出了基于特征融合的目标定位方法和基于深度卷积神经网络的目标识别方法。在目标定位问题上,通过提取HOG特征和LBP特征,串行融合后使用支持向量机作为分类器。实验证明该方法可以对含交通标识的图片进行有效定位,并能够排除不含交通标识的图片干扰。深度卷积神经网络是近年来提出的区别于浅层神经网络的机器学习方法模型,因其优秀的学习能力和应用效果受到广泛重视。介绍了自动编码机、稀疏编码、受限玻尔兹曼机、深度信念网络和卷积神经网络等原理和训练方法,重点介绍了ALex Net和Google Net等深度卷积神经网络模型。根据研究对象和应用场景,提出了针对交通标识识别的深度卷积神经网络模型TSR9L-Net,并建立了相应的训练图像数据库。通过平衡识别率和识别速度,提出一个含9层的轻量级参数数量模型,其中权重层为6层。分别对含7类警告标识和15类禁令标识的样本训练集进行训练,同时对比Le Net-5、Alex Net和TSR9L-Net三种模型的训练效果。其中TSR9L-Net能够在保证准确率的前提下,提升识别速度。GPU硬件平台下,7类标识每批40张识别速度达29.3ms,准确率99.09%;15类标识每批40张识别速度32.0ms,准确率99.29%。无论是识别率还是识别速度,都优于Alex Net。
[Abstract]:Automobile plays a more and more important role in people's life. Safe and smooth driving environment is the ideal state of traffic system. Traffic identification is an important part of intelligent transportation system. It mainly includes two parts: target location and target recognition of traffic signs. A traffic identification method based on feature fusion and deep convolution neural network is proposed. Firstly, the research status of traffic sign recognition at home and abroad is introduced, and the advantages and disadvantages of target location and target recognition methods in previous research are compared. In this paper, a target location method based on feature fusion and a target recognition method based on deep convolution neural network are proposed. In the problem of target location, HOG features and LBP features are extracted. After serial fusion, support vector machine is used as classifier. Experimental results show that this method can effectively locate images with traffic signs. The deep convolution neural network is a machine learning method model proposed in recent years, which is different from the shallow neural network. The principles and training methods of automatic coding machine, sparse coding, constrained Boltzmann machine, depth belief network and convolution neural network are introduced. The deep convolution neural network models such as ALex Net and Google Net are introduced in detail. In this paper, a deep convolution neural network model TSR9L-Netfor traffic identification is proposed, and the corresponding training image database is established. By balancing recognition rate and recognition speed, a 9-layer lightweight parameter quantity model is proposed. The weight layer is six layers. The training sets with 7 warning marks and 15 ban marks are trained, and the training effects of Le Net-5N Alex Net and TSR9L-Net are compared. TSR9L-Net can ensure the accuracy of the training. On the hardware platform of GPU, the recognition speed of each batch of 40 marks is up to 29.3 ms.The accuracy rate of 99.0915 class marks is 32.0ms. the accuracy is 99.290.The recognition rate and speed are better than that of Alex NetNet.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】