整合外觀與邊緣特徵之路面夜間機踏車偵測

在智慧型車輛系統上，障礙物偵測一直是重要的議題，行人因其脆弱性，吸引許多專家進行相關研究。除了行人偵測外，機踏車也是一種重要的障礙物偵測。在過去的文獻中，大多使用整體式機踏車偵測，然而機踏車的寬高比在不同視角的差異，會造成整體式機踏車偵測無法取得好結果。尤其在夜間環境下，光源不如日間明亮，更會造成部分機踏車外觀不清楚。因此本篇論文提出一個在移動式車輛上使用之基於近紅外線光源的夜間路面機踏車偵測方法，在近紅外線影像中會顯示輻射近紅外線以及反射經由近紅外線投射燈投出的紅外線的區域。為了解決機踏車的寬高比、視角差異，本篇論文引入分組式組件的偵測方法，使用兩種影像特徵分組擷取，結合這些特徵在影像上與影像中心的相對位置資訊。除了機踏車偵測外，騎著機踏車的人也需要偵測。因為機踏車駕駛姿勢固定，我們採用梯度方向直方圖當成特徵的機踏車駕駛偵測。然而道路有許多物品外觀與機踏車駕駛相似，如路燈、樹等，因此會產生許多錯誤偵測。因機踏車駕駛與其騎乘之機踏車空間關係十分顯著，可以用來降低錯誤偵測。本篇論文提出空間關係改善機踏車駕駛偵測，使用駕駛與機踏車的質心間位移間大小比例來當成所用之空間關係，分析此空間關係的分佈發現可以呈現高斯分佈，利用高斯分佈的機率公式來計算其機率，最終結果再過濾掉可能性低的組合。因為空間認證效果取決於偵測物體的視窗準確性，而分組式機踏車偵測容易造成視窗的浮動，提出視窗修正方法進行修正原本偵測到的視窗，進而提升空間關係認證效果。

關鍵字

機踏車偵測；機踏車騎士偵測；夜間；近紅外線；空間關係

並列摘要

It is critical to detect bicycles and motorcycles on the road because collision of autos with those light vehicles becomes major cause of on-road accidents nowadays especially in the nighttime. Therefore, a vision-based nighttime on-road bicycle and motorcycle detection method relying on use of a camera and near-infrared lighting mounted on an auto vehicle is proposed in this paper. Generally, the objects will reflect near-infrared lighting. However, some components of the bicycles and the motorcycles absorb most infrared lighting and thus make the bicycles and motorcycles hardly recognizable. To cope with this problem, the aforementioned detection method is part-based, which combines the two kinds of features related to the characteristics of bicycles and motorcycles. Also, the information about the geometric relation among all the parts and the object centroid is learned off-line. Due to high computation load, selection of effective parts with better geometric information is imperative for detection. On the other hand, cyclist is also an important object to detect. We adopt a two-fold strategies, where one detects the cyclist by a holistic-based detector, and second is to establish a spatial relationship model between the cyclist and his/her riding vehicle off line. In particular, the second strategy filters out the wrong detection. The performance of spatial relationship validation depends on the tightness of bounding box. Hence, we propose a bounding box refinement to refine the detection results. To validate the proposed results, several experiments are conducted to show that the developed system is reliable in detecting bicycles and motorcycles on the road in the nighttime.

並列關鍵字

On-road Bicycle and Motorcycle Detection ； Cyclist Detection ； Feature Integration ； Spatial Relationship ； Nighttime

參考文獻

[4] N. Dalal and B. Triggs, "Histograms of oriented gradients for human detection," in International IEEE Conference on Computer Vision and Pattern Recognition, 2005, pp. 886-893.

[6] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan, "Object detection with discriminatively trained part-based models," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 32, pp. 1627-1645, 2010.

[7] B. Leibe, A. Leonardis, and B. Schiele, "Robust object detection with interleaved categorization and segmentation," International journal of computer vision, vol. 77, pp. 259-289, 2008.

[8] A. Opelt, A. Pinz, and A. Zisserman, "Learning an Alphabet of Shape and Appearance for Multi-Class Object Detection," International Journal of Computer Vision, vol. 80, pp. 16-44, 2008.

[9] L. Bourdev and J. Malik, "Poselets: Body part detectors trained using 3d human pose annotations," in International IEEE Conference on Computer Vision, 2009, pp. 1365-1372.

國際替代計量

整合外觀與邊緣特徵之路面夜間機踏車偵測

全文下載

主題瀏覽