基于神经网络学习的车辆信息识别技术研究

发布时间：2018-05-03 22:45

本文选题：车辆信息识别 + 卷积神经网络　；参考：《南京邮电大学》2017年硕士论文

【摘要】：近年来,随着车辆的普及,人们对于交通安全的重视程度越来越高,城市道路监控系统应运而生。并且,随着信息化技术的发展,城市道路监控系统逐渐向着自动化,智能化,信息化的方向发展。然而,由于在现实场景下,监控视频的外在环境非常复杂,因此,传统算法很难应对众多环境因素的影响,在实际应用中无法保证鲁棒性。本文在基于神经网络学习理论的基础上,对车辆信息的识别方法做了深入研究,对车辆检测、车型分类、车标识别三种车辆信息的识别提出了可行的解决方法,并通过实验验证了车辆信息识别方法的有效性。本文主要对车辆信息的识别技术做了如下工作:(1)介绍了车辆检测、车型分类和车标识别的原理和研究现状,提出了车辆信息识别系统的结构框架,并做详细阐述;(2)介绍了卷积神经网络理论,并提出了一种改进的基于Fast-RCNN卷积神经网络的车辆检测算法。通过神经网络的强大拟合能力,应对道路监控视频中多种场景下的的车辆检测问题。通过深度学习网络学习得到车辆的卷积特征图,并通过全连接网络最终输出车辆目标的位置及可信程度。此外,本文还给出了暗光条件下的车辆检测算法,用于提高夜晚视频下的车辆检测正确率。(3)针对传统车型分类算法中的不足,本文提出了一种新型的基于自适应聚类的车型分类算法。创新性地提出了车型主特征和弱标签特征,即通过卷积神经网络获得的车型主特征,以及通过极限学习机(ELM)网络获得的车型弱标签特征。最终,将两种特征按比例进行融合,并通过K-means算法聚类,获得车型的特征中心,从而对后续车辆进行车型判别。由于这种方法可以自适应的获取车型特征,因此可以适用于多种类型的场景。(4)给出了车标识别系统的组成结构。提出了一种“由粗到精”逐步定位车标的检测方法。并重点阐述了极限学习机的数学原理和理论模型,采用了基于局部感知野的极限学习机神经网络进行车标识别。并对特征提取滤波器和网络的学习方法进行了改进,提高网络在多种场景下的算法鲁棒性。(5)对算法进行仿真实验,验证算法的可行性和有效性。对算法的优点和不足进行分析,提出改进方案。
[Abstract]:In recent years, with the popularity of vehicles, people pay more and more attention to traffic safety, urban road monitoring system arises at the historic moment. With the development of information technology, urban road monitoring system is developing towards automation, intelligence and informatization. However, because the external environment of the surveillance video is very complex in the real scene, it is difficult for the traditional algorithm to deal with the influence of many environmental factors, and the robustness can not be guaranteed in the practical application. Based on the learning theory of neural network, this paper makes a deep research on the recognition method of vehicle information, and puts forward a feasible solution to the recognition of three kinds of vehicle information, such as vehicle detection, vehicle model classification and vehicle sign recognition. The validity of vehicle information recognition method is verified by experiments. This paper mainly introduces the principle and research status of vehicle detection, vehicle classification and vehicle sign recognition, and puts forward the structure framework of vehicle information recognition system. The theory of convolution neural network is introduced, and an improved vehicle detection algorithm based on Fast-RCNN convolution neural network is proposed. The strong fitting ability of neural network is used to deal with the vehicle detection problem in various scenes in road surveillance video. The convolution feature map of the vehicle is obtained by learning the network in depth, and the position and credibility of the vehicle target are finally output through the full connected network. In addition, a vehicle detection algorithm under dark light is presented in this paper, which is used to improve the accuracy of vehicle detection in night video. This paper presents a novel vehicle classification algorithm based on adaptive clustering. The main features and weak label features of vehicle type are proposed innovatively, that is, the main feature of vehicle model obtained by convolution neural network and the weak label feature of vehicle model obtained by means of extreme learning machine (ELM) network. Finally, the two features are fused proportionally, and the feature center of the vehicle is obtained by clustering by K-means algorithm, so that the subsequent vehicles can be identified. Because this method can adaptively acquire the vehicle features, it can be used in many kinds of scene. 4) the structure of the vehicle sign recognition system is given. In this paper, a method of detecting vehicle mark from coarse to fine is put forward. The mathematical principle and theoretical model of the LLM are expounded, and the neural network of LLM based on the local perception field is used to identify the vehicle marks. The feature extraction filter and the learning method of the network are improved to improve the robustness of the algorithm in various scenarios. The simulation experiments are carried out to verify the feasibility and effectiveness of the algorithm. The advantages and disadvantages of the algorithm are analyzed and an improved scheme is proposed.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【相似文献】