整合雙眼與單眼視覺技術的行人偵測 ; Pedestrian Detection by Integrating Binocular and Monocular Vision Methods

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/44688

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/44688

Title:	整合雙眼與單眼視覺技術的行人偵測;Pedestrian Detection by Integrating Binocular and Monocular Vision Methods
Authors:	吳怡君;Yi-chun Wu
Contributors:	資訊工程研究所
Keywords:	影像矯正;相機校正;camera calibration;image rectification
Date:	2010-07-26
Issue Date:	2010-12-09 13:52:47 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著經濟的成長，機動車輛愈來愈多，因而交通事故也愈來愈多。有鑑於此，發展車輛輔助安全駕駛的議題，也就愈顯示其重要性。市區的行人防撞是其中一項重要議題。在本研究中，我們提出整合雙眼與單眼視覺技術做行人偵測，並應用於車輛前方上，以避免己車撞及行人。在車輛前方的行人偵測中，我們先將左右影像經由雙相機校正方法，依序透過相機參數校正、扭曲校正、以及影像矯正，得到左右極線彼此平行且對齊的影像。其中，經過影像矯正後的影像與光軸的交點已經改變，所以在此我們將影像矯正後的影像再做一次相機參數校正，得到目前影像與相機的新關係。接著進入視覺技術偵測，首先透過關聯式動態規劃法計算出像差圖；將像差圖經由形態學平滑化消除雜訊；利用v-像差將地面資訊濾除，接著產生連結區塊，將區塊根據在影像上的長度及距離分為長度太小與夠大兩類；長度太小的區塊則根據單眼影像的色彩資訊判斷相鄰且距離差距不大的區塊是否為同一物體，接著與長度夠大的區塊一起經由地面消失線濾除高於地面太多的物體，並判斷區塊的距離及長度是否符合我們定義的長度範圍；最後根據行人的長寬比例框出結果並根據像差值求得該物體與攝影機之距離。In these few decades, the vehicle number is rapidly increasing. In addition to the vehicle number, more factors of road situation, driving environment, and human attention result in a large amount of traffic accidents and casualties. Therefore, it is important to develop real-time automotive driver assistance systems. Pedestrian collision avoidance is one of the important issues. In this study, we propose integrating binocular and monocular vision methods for pedestrian detection, and apply in preceding vehicle detection to avoid collision. In pedestrian detection, we use images of left and right cameras to obtain epipolar lines of left and right images by camera calibration and image rectification. We can obtain the refined images of left and right by aligning the epipolar lines. However, the optical axis of the refined images have been changed because of image rectification, we must obtain the relationship between new images and cameras by camera calibration. Then, we use associated dynamic programming algorithm to obtain disparity map from new images. We reduce noise of disparity map by morphology smoothing, and we filter out the ground by v-disparity. We divide the disparity map into many components by connected component algorithm. We can determine whether the component is pedestrian according to the rules of pedestrian determination and estimate the distance between camera and detected object.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	849	View/Open

社群 sharing

Loading...