基于深度學(xué)習(xí)的交通視頻檢測及車型分類研究
發(fā)布時間：2018-11-18 13:10
【摘要】：隨著汽車保有量的急劇增加,交通問題越來越突出。與此同時,在互聯(lián)網(wǎng)大數(shù)據(jù)時代的背景下,深度學(xué)習(xí)獲得了迅猛發(fā)展,給模式識別任務(wù)帶來了巨大的變革,它還給許多領(lǐng)域提供了一種新的解決方案。因此,將深度學(xué)習(xí)應(yīng)用到解決交通問題已經(jīng)成為一種研究趨勢。本文利用深度學(xué)習(xí)中的卷積神經(jīng)網(wǎng)絡(luò)方法來解決交通視頻中的交通目標檢測及車型分類問題,為智能交通系統(tǒng)提供技術(shù)支持從而緩解交通擁堵等問題。本文主要內(nèi)容如下:首先介紹了深度學(xué)習(xí)的基本模型,主要分為深度置信網(wǎng)絡(luò)、棧式自編碼網(wǎng)絡(luò)和卷積神經(jīng)網(wǎng)絡(luò),主要重點研究了卷積神經(jīng)網(wǎng)絡(luò)的構(gòu)成、卷積神經(jīng)網(wǎng)絡(luò)區(qū)別于傳統(tǒng)神經(jīng)網(wǎng)絡(luò)的特點,以及卷積神經(jīng)網(wǎng)絡(luò)的訓(xùn)練機制。針對利用人工設(shè)計的學(xué)習(xí)特征進行交通目標檢測時,會存在學(xué)習(xí)特征設(shè)計過程繁瑣、適應(yīng)范圍受限制等問題,本文采用卷積神經(jīng)網(wǎng)絡(luò)來自動提取特征。以基于區(qū)域的卷積神經(jīng)網(wǎng)絡(luò)(RCNN)為基礎(chǔ),設(shè)計了交通視頻檢測方案,結(jié)合了Fast RCNN框架和RPN區(qū)域建議網(wǎng)絡(luò)的優(yōu)點。針對交通目標輪廓形狀各異的特點,本文對交通視頻檢測網(wǎng)絡(luò)中的共享卷積網(wǎng)絡(luò)進行了改進,主要是加深了卷積網(wǎng)絡(luò)的深度,從5層卷積加深到13層。在交通訓(xùn)練樣本中取得了較好的效果,交通目標的平均檢測率提升了超過3%。針對已有的車型分類手段只將車輛進行粗略分類,已經(jīng)無法滿足車聯(lián)網(wǎng)對車輛信息需求的問題,本文采用深度殘差神經(jīng)網(wǎng)絡(luò)對車型進行精細型分類,車輛品牌可達64種,車型可達281種。在設(shè)計車型分類網(wǎng)絡(luò)的過程中,分析了常用圖像分類卷積神經(jīng)網(wǎng)絡(luò),并在兩套數(shù)據(jù)集上進行了性能對比,最終選擇了深度殘差網(wǎng)絡(luò)作為車型分類網(wǎng)絡(luò)的主體框架。利用標準車型數(shù)據(jù)集CompCars對車型分類網(wǎng)絡(luò)進行可學(xué)習(xí)參數(shù)微調(diào),訓(xùn)練后的車型分類網(wǎng)絡(luò)的前五準確率在CompCars數(shù)據(jù)集上可達97.3%,在Vehicle ID數(shù)據(jù)集上可達89.4%,驗證了車型分類網(wǎng)絡(luò)的有效性。最后,對本文設(shè)計的基于交通視頻的檢測網(wǎng)絡(luò)和車型分類網(wǎng)絡(luò)分別在圖像和視頻上進行了檢驗。檢測網(wǎng)絡(luò)能在晴天、黑夜、雨天和擁堵等不同狀態(tài)獲得較高的檢測率,在有效視野中車輛檢測率最高可達98.7%,并具有一定的魯棒性。分類網(wǎng)絡(luò)在基于視頻產(chǎn)生的車輛圖像測試集中,獲得了最高達到88%的前五準確率。實驗結(jié)果表明,本文所設(shè)計的檢測網(wǎng)絡(luò)和分類網(wǎng)絡(luò)具有一定的實用價值。
[Abstract]:With the rapid increase of vehicle ownership, traffic problems become more and more prominent. At the same time, under the background of Internet big data era, in-depth learning has developed rapidly, which has brought a great change to the task of pattern recognition. It also provides a new solution in many fields. Therefore, the application of deep learning to solve traffic problems has become a research trend. In this paper, the convolution neural network method in depth learning is used to solve the traffic target detection and vehicle classification problems in traffic video, and to provide technical support for intelligent transportation system to alleviate traffic congestion and so on. The main contents of this paper are as follows: firstly, the basic model of deep learning is introduced, which is divided into three parts: depth confidence network, stack self-coding network and convolutional neural network. Convolutional neural networks are different from traditional neural networks and the training mechanism of convolutional neural networks. In order to solve the problem that the design process of learning features is cumbersome and the scope of adaptation is limited when using artificial design learning features to detect traffic targets, this paper uses convolution neural network to extract features automatically. Based on the area-based convolution neural network (RCNN), a traffic video detection scheme is designed, which combines the advantages of the Fast RCNN framework and the RPN regional recommendation network. In this paper, the shared convolution network in the traffic video detection network is improved, which mainly deepens the depth of the convolutional network, from five layers to 13 layers. Good results were obtained in traffic training samples, and the average detection rate of traffic targets increased by more than 3 percent. In view of the existing vehicle classification methods only rough classification of vehicles, can no longer meet the needs of vehicle information, this paper uses the depth residual neural network for fine classification of vehicle models, vehicle brands can reach 64, There are 281 types of models. In the course of designing the vehicle classification network, the neural network of image classification is analyzed, and the performance of the two sets of data sets is compared. Finally, the depth residual network is chosen as the main frame of the vehicle classification network. The model classification network can be fine-tuned by using the standard model data set (CompCars). The first five accuracy rates of the trained vehicle classification network can reach 97.3 on the CompCars data set and 89.4 on the Vehicle ID data set. The validity of vehicle classification network is verified. Finally, the detection network based on traffic video and vehicle classification network designed in this paper are tested on image and video, respectively. The detection network can obtain high detection rate in different states such as sunny, dark, rainy and congested. In the effective field of vision, the vehicle detection rate can be up to 98.775, and it is robust to a certain extent. In the vehicle image test set based on video, the first five accuracy rates of classification network are up to 88%. The experimental results show that the detection network and classification network designed in this paper have some practical value.
【學(xué)位授予單位】：哈爾濱工業(yè)大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41;TP183
本文編號：2340119
[Abstract]:With the rapid increase of vehicle ownership, traffic problems become more and more prominent. At the same time, under the background of Internet big data era, in-depth learning has developed rapidly, which has brought a great change to the task of pattern recognition. It also provides a new solution in many fields. Therefore, the application of deep learning to solve traffic problems has become a research trend. In this paper, the convolution neural network method in depth learning is used to solve the traffic target detection and vehicle classification problems in traffic video, and to provide technical support for intelligent transportation system to alleviate traffic congestion and so on. The main contents of this paper are as follows: firstly, the basic model of deep learning is introduced, which is divided into three parts: depth confidence network, stack self-coding network and convolutional neural network. Convolutional neural networks are different from traditional neural networks and the training mechanism of convolutional neural networks. In order to solve the problem that the design process of learning features is cumbersome and the scope of adaptation is limited when using artificial design learning features to detect traffic targets, this paper uses convolution neural network to extract features automatically. Based on the area-based convolution neural network (RCNN), a traffic video detection scheme is designed, which combines the advantages of the Fast RCNN framework and the RPN regional recommendation network. In this paper, the shared convolution network in the traffic video detection network is improved, which mainly deepens the depth of the convolutional network, from five layers to 13 layers. Good results were obtained in traffic training samples, and the average detection rate of traffic targets increased by more than 3 percent. In view of the existing vehicle classification methods only rough classification of vehicles, can no longer meet the needs of vehicle information, this paper uses the depth residual neural network for fine classification of vehicle models, vehicle brands can reach 64, There are 281 types of models. In the course of designing the vehicle classification network, the neural network of image classification is analyzed, and the performance of the two sets of data sets is compared. Finally, the depth residual network is chosen as the main frame of the vehicle classification network. The model classification network can be fine-tuned by using the standard model data set (CompCars). The first five accuracy rates of the trained vehicle classification network can reach 97.3 on the CompCars data set and 89.4 on the Vehicle ID data set. The validity of vehicle classification network is verified. Finally, the detection network based on traffic video and vehicle classification network designed in this paper are tested on image and video, respectively. The detection network can obtain high detection rate in different states such as sunny, dark, rainy and congested. In the effective field of vision, the vehicle detection rate can be up to 98.775, and it is robust to a certain extent. In the vehicle image test set based on video, the first five accuracy rates of classification network are up to 88%. The experimental results show that the detection network and classification network designed in this paper have some practical value.
【學(xué)位授予單位】：哈爾濱工業(yè)大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41;TP183
【參考文獻】
相關(guān)期刊論文 前10條
1 杜杰;吳謹;朱磊;;基于區(qū)域特征融合的RGBD顯著目標檢測[J];液晶與顯示;2016年01期
2 李正熙;;中國城市智能交通系統(tǒng)產(chǎn)業(yè)化發(fā)展趨勢[J];自動化博覽;2015年07期
3 趙娜;袁家斌;徐晗;;智能交通系統(tǒng)綜述[J];計算機科學(xué);2014年11期
4 藺宏良;黃曉鵬;;車聯(lián)網(wǎng)技術(shù)研究綜述[J];機電工程;2014年09期
5 賈建英;董安國;;基于聯(lián)合直方圖的運動目標檢測算法[J];計算機工程與應(yīng)用;2016年05期
6 蘇靜;王冬;張菲菲;;車聯(lián)網(wǎng)技術(shù)應(yīng)用綜述[J];物聯(lián)網(wǎng)技術(shù);2014年06期
7 陸化普;李瑞敏;;城市智能交通系統(tǒng)的發(fā)展現(xiàn)狀與趨勢[J];工程研究-跨學(xué)科視野中的工程;2014年01期
8 萬文利;胡加佩;劉學(xué)軍;;基于誤差橢圓的車型識別算法[J];計算機工程;2012年05期
9 任建強;;基于視頻序列的車型識別算法設(shè)計[J];計算機工程;2011年24期
10 秦鐘;;基于圖像不變矩特征和BP神經(jīng)網(wǎng)絡(luò)的車型分類[J];華南理工大學(xué)學(xué)報(自然科學(xué)版);2009年02期
相關(guān)博士學(xué)位論文 前1條
1 胡耀民;基于視頻的車型識別關(guān)鍵技術(shù)研究[D];華南理工大學(xué);2014年
相關(guān)碩士學(xué)位論文 前5條
1 周標;交通監(jiān)控視頻中的車輛檢測技術(shù)研究[D];華南理工大學(xué);2016年
2 張飛云;基于深度學(xué)習(xí)的車輛定位及車型識別研究[D];江蘇大學(xué);2016年
3 張文桂;基于深度學(xué)習(xí)的車輛檢測方法研究[D];華南理工大學(xué);2016年
4 楊康;基于視頻的車輛檢測理論與方法研究[D];長安大學(xué);2013年
5 崔瑩瑩;智能交通中的車型識別研究[D];電子科技大學(xué);2013年
，本文編號：2340119
本文鏈接：http://sikaile.net/kejilunwen/ruanjiangongchenglunwen/2340119.html
最近更新
教材專著
