醫療載具的單眼視覺行人偵測

Monocular-vision Pedestrian Detection for Medical Vehicles

陳卓仁(Gho-Jen Chen)

指導教授 ： 曾定章 ； 鄔蜀威

國立中央大學/工學院/生物醫學工程研究所/碩士(2012年)

摘要

隨著交通工具使用的頻繁與流行，交通事故成為廣受國人注意的安全議題。交通事故中若波及行人，不論對駕駛人或行人都將造成相當嚴重的損傷。本研究中針對醫療載具；例如，救護用車、殘障專用車輛、輪椅、輔助行動工具等。這類載具的駕駛人需要投注更多專注力在駕駛或是本身注意能力、反應能力較一般人弱勢，我們在本研究中提出一個使用單眼視覺的行人偵測系統，並應用於複雜的校園、街道環境中，以避免發生與行人碰撞的事故。
在本論文中，我們以梯度方向分佈圖 (Histograms of oriented gradients, HOG) 作為表達行人特徵的方式，行人分類學習步驟中，我們引入支援向量學習機 (Support Vector Machine, SVM) 作為學習分類器。基於單眼視覺的行人偵測需要產生大量的候選行人視窗供給分類器作判別，然而，HOG 特徵是以累積視窗內像素點各方向的邊強度取得，當候選視窗數量眾多，累積所需得執行時間將會大幅升高，本研究使用行人偵測帶 (pedestrian detection strip, PDS) 作為第一步蒐尋包含行人位置的 ROI 區塊，再引入距離轉換 (distance transform, DT) 模板作第一步的候選行人視窗篩選，通過初步篩選後的視窗再交給 SVM 作最後的判定。
在行人偵測的結果方面，在巷弄間複雜的背景、校園內行人數多的情況下、以及光照不同的情形下，交給 SVM 分類的行人視窗辨識效果都有 94% 以上的正確率，而在一般的情況下影格中出現行人的偵測率也約有 80% 以上，而每個影格中出現 false positive 的偵測錯誤也指有 0.3 個左右。

關鍵字

距離轉換 ； 行人偵測 ； 梯度方向分佈圖

並列摘要

With frequent and popular use of transport, traffic accidents become great issues for people to pay attention. If there were pedestrians injured in an accident, whether drivers or pedestrians will cause very serious loss. In this study, we pay more attention to some medical vehicles, such as ambulance, disabled vehicle, wheelchair and action assistance tools. For those vehicles, the drivers should pay more attention in the driving or the drivers would have weaker reaction ability than a normal person. We proposed a pedestrian detection system using monocular vision. To avoid collision accident, we applied our system to some complex environment, such as campus and street situation.
In this paper, we use Histograms of oriented gradients (HOG) as a way of pedestrian characteristics expression. We introduce the support vector machine (SVM) for pedestrian classification learning. The classifier training has to generate many candidate of pedestrian window. However, HOG features are accumulated by each direction of edge intensity. It will cause the time of process rising while the number of candidate become larger. The research employs the pedestrian detection strip (PDS) as the first seeking-step of interesting region. Then we using distance transform (DT) template as selection function to sift the candidate windows of pedestrian out. Finally we constructed the HOG feature of each window and used the SVM classifier for final judgment. In the results, we have More than 94% of accuracy with SVM classifier of pedestrian detection. Other side, our detection rate of pedestrian is above 80% and the detection error of false positive in each frame is mean about 0.3 / frame.

並列關鍵字

pedestrian detection ； distance transform ； histograms of oriented gradients

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.

[3] Ahonen, T., A. Hadid, and M. Pietikinen, ''Face description with local binary patterns: application to face recognition,'' IEEE Trans. Pattern Analysis and Machine Intelligence, vol.28, no.12, pp.2037-3041, 2006.

[6] Bertozzi, M., A. Broggi, M. Del Rose, M. Felisa, A. Rakotomamonjy, and F. Suard, ''A pedestrian detector using histograms of oriented gradients and a support vector machine classifier,'' in Proc. IEEE Conf. Intelligent Transportation Systems, Seattle, WA., Sep.30-Oct.3, 2007, pp.143-148.

[9] Chen, C. H. and D. L. Yang, "Fast algorithm and its systolic realization for distance transformation," IEE Proc. Computers and Digital Techniques, vol.143, no.3, pp.168-173, 1996.

[10] Cheng, Y., ''Mean shift, mode seeking, and clustering,'' IEEE Trans. Pattern Analysis and Machine Intelligence, vol.17, no.8, pp.790-799, 1995.