結合 AdaBoost與 SVM 分類器的單眼視覺行人偵測

Combining AdaBoost and SVM classifiers for monocular-vision pedestrian detection

蔡育程(Yu-cheng Cai)

指導教授 ： 曾定章

國立中央大學/資訊電機學院/資訊工程學系/碩士(2014年)

摘要

隨著民眾車輛的增加，交通事故案件也跟著增多；交通事故已經成為影響大眾生命安全的一個重大問題。由於駕駛人的分心或疏忽而碰撞路上行人時有所聞；因此，我們在本研究中提出了一個使用單眼電腦視覺的行人偵測系統，應用於一般街道巷弄和校園中，作為警示駕駛人的行車輔助系統，以避免行人碰撞的交通事故。
我們所提的行人偵測系統主要分為三個步驟：一、使用行人偵測帶 (pedestrian detection strip, PDS) 計算影像中前景物的位置，定義出候選物件視窗；二、對各個視窗比對行人輪廓模板，找出疑似有行人物件的視窗做後續的驗證，以減少大量分類的計算；三、使用梯度方向分佈圖 (histograms of oriented gradients, HOG) 描述行人特徵，並在分類決策上，使用支援向量機 (support vector machine, SVM) 針對行人樣本中的九個區域各別學習訓練做為分類器。最後再使用 AdaBoost 組合這些分類器，分別學習權重，作為系統最後辨識行人的依據。各區域學習訓練的分類器可以減少因為視線角度與行人部分遮蔽的問題，可以針對行人各個輪廓各別判定，並透過 AdaBoost 來整合分類器的結果，使得整體分類能力有所提昇。
在實驗分析中，我們擷取實際一般街道與校園中的行車影片；其中包含多種日間時段與日照情形的影像。在物件偵測的步驟中，透過行人偵測帶的擷取前景物位置可達到99.6%的行人偵測率。輪廓模板比對可初步篩選掉70%以上的非行人視窗，而加速後續的AdaBoost-SVM確認程序。在使用相同的 HOG 特徵之下，本研究所提出的 AdaBoost 結合 SVM 在各種偵測環境下約有87%的偵測率與4%的誤判率；而僅使用SVM 的偵測率只有83%偵測率與7%誤判率；可說明本研究使用的分類器架構較能夠因應環境因素影響與外形多變的行人。最後系統在一般電腦運算上每秒約能執行20張影像的處理速度。

關鍵字

自適應增強 ； 支援向量機 ； 梯度方向分佈圖 ； 行人偵測帶 ； 模板比對

並列摘要

With the increase of public vehicles, traffic accidents increase also followed the case. Traffic accidents have become a major problem affecting the safety of the public, because the driver's distraction or negligence collision pedestrians heard. Therefore, we propose a computer using a monocular vision pedestrian detection system in the study, applied to the general campus streets and alleys, using driving assistance systems as a reminder to driver, in order to avoid pedestrian collision accidents.
Our propose pedestrian detection system is divided into three steps: First, the use of pedestrian detection strip (PDS) to calculate the position of the image in the foreground objects, define the candidate object window. Second, each window would template matching for pedestrian silhouette to identify suspected pedestrian objects window to do the follow-up verification, in order to reduce the computational lot of classification. Third, using the histograms of oriented gradients (HOG) describe characteristics of pedestrians, and on the classification decision, the use of support vector machines (SVM) as a classifier for pedestrian samples nine regional individual learning and training. Finally, the combinations of these classifiers using AdaBoost were learning the weights, as a system based on the final identification of pedestrians. Training of regional learning classifiers can reduce the problem because the line of sight angle and pedestrians partially obscured can determine for each individual contour pedestrians to integrate through AdaBoost classifier results, making the overall classification ability has improved.
In the experimental analysis, we retrieve the actual streets and general campus traffic movie, which contains images with a variety of daytime sunshine circumstances. At step in object detection, object position by capturing of PDS can be achieved with detection rate of 99.6%. Template matching can filter out more than 70% of non-pedestrian window, and accelerate the subsequent AdaBoost-SVM confirmation process. Using the same HOG features, AdaBoost with multiple SVM proposed in this study is about 87% detection rate and false positive rate of 4% is detected in a variety of environments, and compare single SVM detection rate of only 83% to detect rate and false positive rate of 7%, can be explained classification framework used in this study can be compared with the response to environmental factors that affect the shape changing pedestrians. Finally, the system in general can perform computing approximately 20 images per second.

並列關鍵字

AdaBoost ； SVM ； HOG ； PDS ； Template Matching

參考文獻

[1] Agarwal, S., A. Awan, and D. Roth, ''Learning to detect objects in images via a sparse, part-based representation,'' IEEE Trans. Pattern Analysis and Machine Intelligence, vol.26, no.11, pp.1475-1490, 2004.

[2] Ahonen, T., A. Hadid, and M. Pietikainen, ''Face description with local binary patterns: application to face recognition,'' IEEE Trans. Pattern Analysis and Machine Intelligence, vol.28, no.12, pp.2037-2041, 2006.

[3] Alonso, I. P., D. F. Llorca, and M. Á. Sotelo, ''Combination of feature extraction methods for SVM pedestrian detection,'' IEEE Trans. Intelligent Transportation System, vol.8, no.2, pp.292-307, 2007.

[7] Bertozzi, M., A. Broggi, A. Fascioli, T. Graf, and M.-M. Meinecke, ''Pedestrian detection for driver assistance using multiresolution infrared vision,'' IEEE Trans. Vehicular Technology, vol.53, no.6, pp.1666-1678, 2004.

[9] Broggi, A., R. I. Fedriga, and A. Tagliati, ''Pedestrian detection on moving vehicle: an investigation about near infra-red images,'' in Proc. IEEE Conf. Intelligent Vehicles Symp., Tokyo, Japan, Jun.13-15, 2006, pp.431-436.