可適應亮度與距離變化的盲點區域車輛偵測技術;Light- and distance-adaptive vehicle detection in blind-spot areas

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/72017

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/72017

题名:	可適應亮度與距離變化的盲點區域車輛偵測技術;Light- and distance-adaptive vehicle detection in blind-spot areas
作者:	徐震濤;Hsu,Chen-Tao
贡献者:	資訊工程學系
关键词:	先進駕駛輔助系統;盲點偵測;光流;車底陰影;亮度適應性;距離適應性;Advanced driver assistance system;blind spot detection;optical flow;underneath shadow;light adaptive;distance adaptive
日期:	2016-07-01
上传时间:	2016-10-13 14:21:53 (UTC+8)
出版者:	國立中央大學
摘要:	當一位駕駛人行駛在道路上要變換車道時，駕駛人須轉頭掃視側後方觀察在側邊車道是否有接近的來車。然而，以上行為所的視野範圍是有限的，會有一個不可見的盲點區域存在。為避免在變換車道時可能發生的交通事故，我們在此提出一個車道變換輔助系統以輔助駕駛人注意左右來車。在本系統中，將兩個相機裝在己車兩邊後照鏡下方來偵測接近車輛。我們以來車的光流特徵、邊點特徵，及車底陰影等動靜態特徵來偵測來車。在此視覺偵測中，有兩個主要的問題會影響偵測結果，一個是環境明暗度的問題，另一個是遠近不同造成來車大小與速度不同的透視投影效應問題。在本研究中，我們提出兩個適應性的方法：亮度適應性及距離適應性方法來克服這兩個問題。在戶外環境中，駕駛人必須面對不同的天候與環境變化，這些不同的天候與環境會造成影像內容有明暗的變異而使得車輛偵測變得更困難。本研究使用一個二維亮度/梯度分佈圖(2D brightness/gradient magnitude histogram)在不同的亮度變化下，適應地判斷出邊點、陰影、路面標線，及其他不同成份。另一方面，當側邊車輛在真實空間以更高的速度接近己車時，在影像平面上會因透視投影效應，來車速度與大小會增加。在本研究中，我們也提出一個距離適應性的方法來補償影像平面中水平方向的光流長度。經由距離適應性補償後，水平方向上的光流長度則不受距離影響。除了以上亮度與距離適應性的問題外，此盲點偵測系統也克服了路標、樹蔭的誤判，及側邊等速車的偵測不到的問題。本方法包含四個階段：i. 二維數量分佈圖估算亮度適應性門檻值、ii.光流估算、iii. 靜態特徵擷取，及iv. 結合動靜態偵測的決策法則。在實驗中，我們以14段測試影片中的6842張影像，來評估系統的表現度。這些影片中，有六種不同亮度情形。首先，我們比較了有無亮度適應性的表現度。在無亮度適應性的方法中，我們將三種門檻值(車底陰影、路面標線，以及邊點) 27組不同給定值，並將此27組值分別應用於五種方法：僅靜態(S)、僅動態(M)、動靜皆有(S&M)、動靜擇一(SorM)，以及不包含亮度適應性之動靜結合(SM)等。比較此五種方法後發現，我們需要一個亮度適應性的方法用以動態調整三種門檻值。再加入亮度適應性的方法後，動靜結合配合亮度適應性的方法準確率可達91.84%的正確率、7.12%的誤判率，以及1.04漏失率。動靜結合有亮度適應性準確率較無亮度適應性增加16.68%。在距離適應性功能上，首先，我們證明了經過補償後，水平方向光流可不受距離的影響。其次，我們發展了一個距離適應性的方法來偵測側邊車輛，距離適應性方法可達93.88%準確率、5.36%誤判率，以及0.76%漏失率。 ;A driver wants to change lane when driving on a road, he must glance the rearview and outside mirrors of his vehicle and turn his head to scan the possible approaching vehicles on the side lanes. However, the field of view by the above behavior is limited; there is a blind spot area invisible. To avoid the possible traffic accident during lane change, we here propose a lane change assistance system to assist drivers changing lane. In this system, two cameras are mounted under outside mirrors of the host vehicle to capture rear-side-view images for detecting approaching side vehicles. In this application of visual detection, there are two main problems influencing the detection results, they are various light conditions and perspective projection effect. In this study, we present two adaptive methods to overcome the two problems. Drivers have to face various weather and environment conditions when driving. The different weather conditions cause various light conditions, and the various light conditions may influence the detection results. The proposed method uses a 2-D intensity/gradient histogram to adaptively judge the edge, shadow, lane marks, and other scene components in various light conditions. On the other hand, when a side vehicle runs toward to the host vehicle with constant speed, the size and speed of the side vehicle appear increasing in images due to perspective projection effect. In this study, we propose a distance-adaptive method to compensate the horizontal optical-flow vectors in images. After the compensation, the horizontal optical-flow vectors are invariant to distance. Besides the above light and distance variant problems, the blind-spot vehicle detection also encounters the problems of false detection on lane marks and tree shadow on ground, and loss detection on similar-speed side vehicles. Thus, both static and motion features are adopted and sophisticatedly judged to detect side vehicles in this study. The proposed system consists of four stages: estimation of light-adaptive threshold values with a 2-D intensity/gradient histogram, multiresolution optical flow estimation with distance-adaptive compensation, static feature detection, and detection decision based on the static and motion features. In experiments 6842 images in 14 side-vehicle videos were tested to evaluate the performance of the proposed systems. These videos were captured from six kinds of light conditions. Without light-adaptive function, we set 27 groups given values to three kinds of threshold values: underneath shadow, lane marking, and edge points. And also, we apply the 27 groups given value to five detection methods: only static (S), only motion (M), static and motion (S&M), static or motion (SorM), and mutually combing static and motion features without light-adaptive method (SM). Comparing the results of five detection methods, we find that we need the light-adaptive method to adjust the three kinds of threshold dynamically. After applying light-adaptive detection method, the accuracy of mutually combining static and motion features with light-adaptive method achieves 91.84% detection rate, 7.12% false alarm rate, and 1.04% missing rate. The accuracy of mutually combining static and motion features with light-adaptive method improve 16.68% than without light-adaptive method. The missing rate with light-adaptive is lower 19.82% than without light-adaptive method. With distance-adaptive function, first, we prove that the horizontal optical-flow vectors are invariant to distance after compensation. Second, we develop a distance-adaptive detection method to detect side vehicle. The accuracy of distance-adaptive method achieves 93.88% detection rate, 5.36% false alarm rate, and 0.76% missing rate.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	229	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....