本文選題：AdaBoost級聯(lián)分類器　切入點：模板匹配　出處：《東北大學》2014年碩士論文

【摘要】：隨著汽車保有量的增加,我國交通安全問題日益突出,由駕駛員疲勞駕駛造成的交通事故越來越多,現(xiàn)已成為交通事故發(fā)生的主要因素之一。由此可見,研究并實現(xiàn)疲勞檢測相關算法對預防交通事故的發(fā)生有著重大的現(xiàn)實意義。本文對國內外有關駕駛員疲勞檢測的相關技術進行了系統(tǒng)分析,最終選取AdaBoost定位算法和模板匹配跟蹤算法,并做了進一步研究。在定位階段,首先對原圖像進行預處理,主要包括理想光照條件下的圖像灰度化、直方圖均衡化與非理想光照條件下的圖像亮度對比度變換,然后應用人臉AdaBoost級聯(lián)分類器進行人臉定位,并對檢測到的所有的人臉進行判斷,得出真正的駕駛員臉部范圍,進而在此范圍內應用人眼AdaBoost級聯(lián)分類器進行人眼定位,并通過自適應閾值判斷得出正確的人眼,從而最終得到人眼跟蹤所用的模板;在跟蹤階段,根據(jù)連續(xù)兩幀人眼在水平方向和垂直方向的位移預測出人眼區(qū)域,并應用模板匹配對人眼進行實時跟蹤,但傳統(tǒng)的模板匹配算法常常由于累積誤差和人眼的眨動導致跟蹤丟失,所以本文提出在模板匹配的基礎上進行人眼輪廓提取,并根據(jù)人眼輪廓對人眼模板進行更新,從而解決了跟蹤丟失的問題；在疲勞評測階段,根據(jù)提取的人眼狀態(tài)參數(shù)統(tǒng)計眨眼頻率和計算PERCLOS值,最終得到人眼在不同時刻的特征,根據(jù)眼睛狀態(tài)判斷駕駛員疲勞程度。為了避免目標跟蹤丟失,保證跟蹤的準確性和實時性,跟蹤階段還需對人眼進行重定位判斷,在跟蹤出錯或丟失時及時的進行重定位。本文在PC機上采用C#編程語言,使用VS2008開發(fā)環(huán)境并基于OpenCV計算機視覺庫仿真實現(xiàn)了駕駛員疲勞評測算法。對理想條件下不同光照、不同速度、不同旋轉角度和非理想條件下不同顛簸程度、光照突變等情況的實時性和準確性進行測試。根據(jù)測試結果分析可知,在非極端情況下對人臉、眼睛定位與眼睛跟蹤算法在各個階段均能實時準確地實現(xiàn),而在極端情況下實驗結果雖有改進但是效果不是特別明顯。駕駛員通常駕車環(huán)境都是出于非極端條件下,從而驗證了本文所用的算法適用于駕駛員疲勞檢測。
[Abstract]:With the increase of vehicle ownership, traffic safety problems become more and more prominent in China. More and more traffic accidents caused by drivers' fatigue driving have become one of the main factors of traffic accidents.Therefore, it is of great practical significance to study and implement fatigue detection algorithms to prevent traffic accidents.In this paper, the related techniques of driver fatigue detection at home and abroad are systematically analyzed. Finally, AdaBoost location algorithm and template matching tracking algorithm are selected, and further research is done.In the localization stage, the original image is preprocessed, which includes image grayness under ideal illumination, histogram equalization and image brightness contrast transformation under non-ideal illumination.Then face location is carried out by using the face AdaBoost cascade classifier, and all the faces detected are judged, and the real face range of the driver is obtained, and then the human eye AdaBoost cascade classifier is used to locate the human eye in this range.The correct human eye is obtained by adaptive threshold judgment, and the template for eye tracking is finally obtained. In the tracking stage, the human eye region is predicted according to the horizontal and vertical displacement of the human eye in two successive frames.Template matching is used to track human eyes in real time, but the traditional template matching algorithms often lose track due to accumulated errors and blink of human eyes, so this paper proposes to extract human eye contour on the basis of template matching.According to the human eye profile, the human eye template is updated to solve the problem of tracking loss, and in the fatigue evaluation stage, the blink frequency and the PERCLOS value are calculated according to the extracted human eye state parameters, and the characteristics of the human eye at different times are finally obtained.Judge the driver's fatigue according to the state of the eye.In order to avoid the loss of target tracking and ensure the accuracy and real-time of tracking, it is necessary to reposition the human eyes in the tracking stage, and to relocate in time when the tracking is wrong or lost.In this paper, we use C # programming language on PC, use VS2008 development environment and realize driver fatigue evaluation algorithm based on OpenCV computer vision library simulation.The real time and accuracy of different illumination, different speed, different rotation angle, different turbulence degree and illumination sudden change in ideal condition were tested.According to the analysis of the test results, the eye localization and eye tracking algorithms can be realized in real time and accurately in the non-extreme cases, but in extreme cases the experimental results are improved, but the effect is not particularly obvious.Drivers usually drive under non-extreme conditions, which verifies that the algorithm proposed in this paper is suitable for driver fatigue detection.
【學位授予單位】：東北大學
【學位級別】：碩士
【學位授予年份】：2014
【分類號】：U495;TP391.41
，

本文編號：1717559