不受內部類影響之行人偵測

Abstract

行人偵測在智慧型車輛系統上一直都是重要的議題。在過去的文獻中，採用梯度方向直方圖為特徵的行人偵測是最為人所知且成功的。雖然如此，它仍然存在一些弱點，梯度方向直方圖在平坦和雜亂的影像上準確率容易受到影響，錯誤偵測很容易出現在這些區域上。為了加強最後偵測的結果，在本篇論文中提出了一個以基於影像強度比較的特徵，稱為Local Oriented Pattern (LOP) 去補足平坦與雜亂影像的判斷。LOP主要的包含兩個資訊：為影像材質與表現強度。藉由將影像中的每個像素轉換成圖形與強度，去統計小區域影像中的變化與分布。在行人偵測實驗中，可以看到LOP相對於梯度方向是更精簡與有效率的。另外在偵測器學習上面，本研究也提出一個新的學習策略，在不改變偵測器效能的情況下，可以使用更少的記憶體與時間去完成訓練一個偵測器，對於後續的實驗與研究上都有實質的幫助。 行人偵測上另一個重要的議題是行人的外觀變化很大，常見的行人姿勢包括行走、跑步與騎乘動作，這些由於姿勢的變化造成外觀的不同會影響到學習演算的成效，要找到共有的特徵會更為困難。為了要達到更好的效能，提出一個樣本為基礎的分類的演算，將訓練的資料的行人動作做分類，資料越小越紮實將有助於演算法的學習。採用誤差修正碼 (Error Correcting Output Code, ECOC) 為去訓練出最後多類別的最終偵測器，ECOC可以結合多的線性偵測器去達到非線性的決策曲面，進而提升學習的結果。最後結合所提出的LOP與ECOC搭配支持向量機去建構最終的分類器。

Pedestrian detection is an important part of intelligent transportation systems. In the literature, the use of Histogram of Oriented Gradients (HOG) feature for pedestrian detection is well known for its good performance, but there are still some false detections appearing in the cases with flat area or clustered background. To deal with the false positive problems, in this research work we develop a new feature which is based on local intensity comparison, called Local Oriented Pattern (LOP). The idea of LOP is to encode the saliency of image and textural information of local area, which describing how different the pixel intensities are distributed within a region. Each pixel is represented as a pattern and its magnitude. It is shown that the special characteristics of LOP feature are “small” and “efficiency” relative to HOG. We also present a training scheme that can be applied to a huge database for training a detector. Such training scheme can reduce the number of hard samples during bootstrap training. Using our training scheme can save the memory as well as the training time for training a detector. Another issue of pedestrian detection is that the human posture changes when the person in different states of walking, running or riding. In addition, different viewpoints caused by moving camera also produce different human appearances. To achieve higher detection rate in the intra-class variation problem, we propose an exemplar-based clustering algorithm to separate the training data into small and compact set. Moreover, the employed Error Correcting Output Code (ECOC) method constructs a nonlinear classification boundary that can discriminate the pedestrian from negative samples. We use ECOC to train multiple base classifiers with LOP feature and linear Support Vector Machine (SVM).