基于盲源分離的車輛檢測與分類技術(shù)研究
發(fā)布時間：2018-11-02 15:31
【摘要】：智能交通系統(tǒng)近些年得到快速發(fā)展,其關(guān)鍵技術(shù)是運動車輛的檢測與分類�；趬毫鞲衅鞯葌鹘y(tǒng)的檢測與分類技術(shù)存在安裝、維護成本高昂等缺點,而基于視頻圖像的車輛檢測與分類技術(shù)能克服上述缺點。現(xiàn)有基于視頻的分類技術(shù)主要歸結(jié)為四類:車牌識別法、基于車輛幾何特征分類法、基于車輛外形輪廊匹配法與基于PCA和LDA的代數(shù)特征法。車牌識別法受車牌信息庫與遮擋車牌情況的限制;幾何特征(長,寬,高等)分類準確率低下,只適用于粗略分類;基于提取外形輪廊進行模板匹配的方法,實際應(yīng)用中往往難以提取到完整輪廊,導(dǎo)致分類準確率不高,并且只適用于輪廊信息明顯的側(cè)面視角監(jiān)控視頻;基于圖像低階信息的代數(shù)特征在應(yīng)用上簡單,且不受監(jiān)控視頻視角的影響,但同樣存在分類準確率不高的問題。車輛分類是交通監(jiān)控系統(tǒng)的關(guān)鍵,本文將盲源分離算法引入到車輛分類問題的研究中,利用ICA能夠去除信號高階統(tǒng)計相關(guān)性的特點進行車輛圖像的特征提取。通過實驗分析ICA的兩種結(jié)構(gòu)所提取得到的車輛特征在分類性能上的優(yōu)劣,實驗驗證ICA模型提取的車輛特征在分類性能上要優(yōu)于傳統(tǒng)的PCA與LDA算法。在此基礎(chǔ)上,提取車輛的幾何特征(長與寬的和)并根據(jù)車輛在圖像中的位置對該特征進行修正,并將該特征用于對車型大小進行預(yù)分類,結(jié)合代數(shù)特征,構(gòu)建二次分類系統(tǒng)�；趲缀翁卣髋c代數(shù)特征的二次分類系統(tǒng),能進一步提升車輛分類的準確率。運動車輛的檢測是分類的基礎(chǔ),本文針對道路監(jiān)控視頻特點,對三幀差法進行改進。針對運動車輛表面顏色比較均勻一致或者車身顏色與背景路面顏色相近的情況,傳統(tǒng)三幀差法無法提取到完整的車輛圖像區(qū)域,改進的三幀差法對這些情況具有更好的適應(yīng)性并保持了運行速度快的優(yōu)點,能滿足實時性的要求。在運動檢測中,基于正面視角的道路監(jiān)控視頻,前后車距較小時檢測會誤將兩輛車當作一輛車,提出使用形態(tài)學(xué)腐蝕算子進行粘連車輛的分離,通過定位兩輛車的中心進行車輛圖像分割。本文初步設(shè)計并實現(xiàn)了一個基于OpenCV與VS2010的道路監(jiān)控系統(tǒng),該系統(tǒng)可以實現(xiàn)運動車輛的檢測、跟蹤、計數(shù)與分類功能。
[Abstract]:Intelligent Transportation system (its) has been developing rapidly in recent years. The key technology of its is the detection and classification of moving vehicles. The traditional detection and classification technology based on pressure sensor has some disadvantages such as installation high maintenance cost and so on. However vehicle detection and classification technology based on video image can overcome these shortcomings. The existing video classification techniques can be divided into four categories: license plate recognition method, vehicle geometric feature classification method, vehicle profile wheel-corridor matching method and algebraic feature method based on PCA and LDA. The method of license plate recognition is limited by the information base of license plate and occlusion of license plate, the accuracy of geometric feature (length, width, high etc.) is low, so it is only suitable for rough classification. Based on the method of template matching, it is difficult to extract the complete wheel corridor in practical application, which leads to the low classification accuracy and is only suitable for the profile visual angle surveillance video with obvious profile information. The algebraic feature based on low order information of image is simple in application and not affected by the visual angle of surveillance video, but it also has the problem of low classification accuracy. Vehicle classification is the key of traffic monitoring system. In this paper, blind source separation algorithm is introduced into the study of vehicle classification problem, and the feature extraction of vehicle image can be carried out by using ICA which can remove the high order statistical correlation of signal. The classification performance of the vehicle features extracted from the two structures of ICA is analyzed experimentally. The experimental results show that the vehicle features extracted by the ICA model are superior to the traditional PCA and LDA algorithms in classification performance. On this basis, the geometric feature (the sum of length and width) of the vehicle is extracted and modified according to the position of the vehicle in the image. The feature is used to pre-classify the vehicle size and the quadratic classification system is constructed by combining the algebraic features. The quadratic classification system based on geometric and algebraic features can further improve the accuracy of vehicle classification. The detection of moving vehicles is the basis of classification. According to the characteristics of road surveillance video, the three frame difference method is improved in this paper. When the surface color of moving vehicle is uniform or the color of the body is similar to that of the background road, the traditional three-frame difference method can not extract the complete vehicle image area. The improved three-frame difference method has better adaptability to these situations and keeps the advantage of fast running speed, which can meet the requirement of real-time. In the motion detection, the road surveillance video based on the positive angle of view, when the distance between the two vehicles is small, will be mistaken as a vehicle, and the morphological corrosion operator is used to separate the adhesion vehicle. The vehicle image is segmented by locating the center of the two vehicles. In this paper, a road monitoring system based on OpenCV and VS2010 is designed and implemented. The system can detect, track, count and classify moving vehicles.
【學(xué)位授予單位】：電子科技大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2015
【分類號】：U495;TP391.41
[Abstract]:Intelligent Transportation system (its) has been developing rapidly in recent years. The key technology of its is the detection and classification of moving vehicles. The traditional detection and classification technology based on pressure sensor has some disadvantages such as installation high maintenance cost and so on. However vehicle detection and classification technology based on video image can overcome these shortcomings. The existing video classification techniques can be divided into four categories: license plate recognition method, vehicle geometric feature classification method, vehicle profile wheel-corridor matching method and algebraic feature method based on PCA and LDA. The method of license plate recognition is limited by the information base of license plate and occlusion of license plate, the accuracy of geometric feature (length, width, high etc.) is low, so it is only suitable for rough classification. Based on the method of template matching, it is difficult to extract the complete wheel corridor in practical application, which leads to the low classification accuracy and is only suitable for the profile visual angle surveillance video with obvious profile information. The algebraic feature based on low order information of image is simple in application and not affected by the visual angle of surveillance video, but it also has the problem of low classification accuracy. Vehicle classification is the key of traffic monitoring system. In this paper, blind source separation algorithm is introduced into the study of vehicle classification problem, and the feature extraction of vehicle image can be carried out by using ICA which can remove the high order statistical correlation of signal. The classification performance of the vehicle features extracted from the two structures of ICA is analyzed experimentally. The experimental results show that the vehicle features extracted by the ICA model are superior to the traditional PCA and LDA algorithms in classification performance. On this basis, the geometric feature (the sum of length and width) of the vehicle is extracted and modified according to the position of the vehicle in the image. The feature is used to pre-classify the vehicle size and the quadratic classification system is constructed by combining the algebraic features. The quadratic classification system based on geometric and algebraic features can further improve the accuracy of vehicle classification. The detection of moving vehicles is the basis of classification. According to the characteristics of road surveillance video, the three frame difference method is improved in this paper. When the surface color of moving vehicle is uniform or the color of the body is similar to that of the background road, the traditional three-frame difference method can not extract the complete vehicle image area. The improved three-frame difference method has better adaptability to these situations and keeps the advantage of fast running speed, which can meet the requirement of real-time. In the motion detection, the road surveillance video based on the positive angle of view, when the distance between the two vehicles is small, will be mistaken as a vehicle, and the morphological corrosion operator is used to separate the adhesion vehicle. The vehicle image is segmented by locating the center of the two vehicles. In this paper, a road monitoring system based on OpenCV and VS2010 is designed and implemented. The system can detect, track, count and classify moving vehicles.
【學(xué)位授予單位】：電子科技大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2015
【分類號】：U495;TP391.41
【相似文獻】
相關(guān)期刊論文 前10條
1 陳錫明;黃碩翼;;盲源分離綜述——問題、原理和方法[J];電子信息對抗技術(shù);2008年02期
2 王曉偉;石林鎖;;一種變速率非線性盲源分離算法[J];機械科學(xué)與技術(shù);2012年06期
3 董天寶;楊景曙;;稀疏盲源分離快速算法[J];火力與指揮控制;2012年07期
4 張艷萍;李杰;;回波對消中盲源分離算法的研究與仿真[J];南京信息工程大學(xué)學(xué)報(自然科學(xué)版);2013年06期
5 彭晗,周元建;一種基于信息最大準則的盲源分離新算法[J];湖南工程學(xué)院學(xué)報(自然科學(xué)版);2001年Z1期
6 劉涵,劉丁,劉筱琰;基于高階統(tǒng)計量的自適應(yīng)盲源分離算法[J];西安理工大學(xué)學(xué)報;2002年02期
7 郭潔;沈連豐;宋鐵成;葉芝慧;;基于盲源分離的無線視頻通信研究與仿真[J];東南大學(xué)學(xué)報(自然科學(xué)版);2007年01期
8 張文愛;李喜林;;基于模擬退火思想和粒子群算法的盲源分離[J];科學(xué)技術(shù)與工程;2007年13期
9 馮健;付興龍;劉浩達;李，
本文編號：2306194
本文鏈接：http://sikaile.net/kejilunwen/daoluqiaoliang/2306194.html
