FAST PEDESTRAIN DETECTION WITH CASCADE CLASSIFIERS Ning Zhang, Qixiang Ye, Jianbin Jiao 1 Graduate University of Chinese Academy of Science ABSTRACT In this paper, we propose a method for fast pedestrian detection in images/videos. Multi-scale orientated (MSO) features are proposed to represent coarse pedestrian contour, on which Adaboost classifiers are trained for pedestrian coarse location. In the fine detection, histogram of oriented gradient (HOG) features and SVM classifiers are employed to precisely classify pedestrians and non-pedestrians. The coarse-to-fine scheme can bring out not only a higher speed but also the elimination of smooth image regions that are prone to be falsely detection as positives by strong classifiers. The strong classifier SVM in the fine detection make the detection robust to variance of pedestrian pattern. Experiments validates the proposed method. proposed framework is a classical object detection method for small training set. Despite of the success of the above methods, the employed features are not very competent since these features are sensitive to the object's grey variance, arisen from object appearance, illumination and deformation. In[7], Dalal et al. proposed histogram of oriented gradient (HOG) features for pedestrian representation. They trained a SVM classifier on the features to detect pedestrians in video frames. HOG features are the orientation statistics on image block and have good tolerance to object grey variance and local deformation. We follow their method to do fine pedestrian detection in this paper. In[8][9], the authors proposed the cortex features for object representation using the multi-scale characteristics of Gabor filters. Good performances as these features are, their extraction is extremely time consuming, which is caused by filter operations. In[10], the authors use image co-variance as the feature and Riemannian Manifolds for pedestrian detection. The feature captures the pattern by its global geometry information, and has good tolerance to both grey and local contour variance. In[11], Mu et al. explored Local Binary Pattern (LBP) features' good tolerance to color variance in album images. In[12], Bilattice-based Logical Reasoning is proposed for pedestrian detection considering some deformation and even partial occlusion. By reviewing the existing works, we believe that the pattern variance and computing efficiency problems are still not considered comprehensively. In[13], Munder et al. compare the performance of pedestrian classification with different features and classifiers. They find that local receptive field features can do a better job in representing pedestrians and the combination of SVM/Adaboost classifiers also outperforms the other classifiers. This inspires us to develop the coarseto-fine detection scheme. We first select square image blocks of different sizes as the basic units for contours. Multi-scale orientated (MSO) features on the selected units are
then extracted for Adaboost classifiers with coarse features. The coarse features can be calculated by an"integral image" method[5] with a high calculation speed. Then, we extract HOG features and train a SVM classifier for fine detection. In the detection procedure, Adaboost classifiers are set as the cascades in front, followed by an
Index Terms-Pedestrian detection, Adaboost classifiers, SVM classifier, histogram of oriented gradient features, Multi-scale orientated features 1. INTRODUCTION Detecting pedestrians in images/videos has been being a hot research topic because of its variety of applications, such as visual surveillance, robot vision, perceptual interfaces[1-4], etc. Although it has been widely investigated, it is still challenged by pattern variance arose from clothes and deformation, etc. In the early years, pedestrian detection is mostly based on background subtraction and region shape analysis[2]. In most cases, these methods are suitable for videos with static background; however, they are not competent for moving background or a static image. Most recent works use image features for pedestrian detection. The most representative method is based on Haarlike features and cascade classifiers, e.g. Adaboost, which can obtain extremely good results given large training sets. Additional advantage of this method is its high speed brought out by the coarse-to-fine scheme and fast feature extraction algorithm on"integral image"[5]. In[6], Mohan et al. proposed a pedestrian detection method by learning component classifiers with Haar-like features and the SVM classifier. They are the forerunners on exploring the SVM's advantage on modeling variance of pedestrian pattern. The 1
Corresponding author. Tel.:+86 10 88256968. E-mail: jiaojb@gucas.ac.cn
978-1-4244-5076-3/09/$26.00©2009 IEEE
550

Detection using a Boosted Cascade of Simple ...cascade classifiers Discard the background regions ...applications for a extremely fast face detector ?...
Detection using a Boosted Cascade of Simple ...cascade classifiers Discard the background regions ...applications for a extremely fast face detector ?...
热门文档