%0 Journal Article
%T 一种基于特征金字塔的快速行人检测方法<br>A fast pedestrian detection algorithm based on feature pyramid
%A 王世芳
%A 徐琨
%A 陈明瑶
%J 长安大学学报(自然科学版)
%D 2018
%X 针对由于背景变化，行人尺度的不确定性以及遮挡等因素的存在下，如何提高检测精度和速度问题。基于相邻尺度通道特征的可预测性，提出一种基于特征金字塔的快速行人检测方法。首先，计算关键尺度下的聚合通道特征，该特征由3个LUV通道特征、1个局部量纲一化的梯度幅值通道特征和6个梯度方向直方图（HOG）通道特征构成，充分反映了图像的梯度信息和颜色信息。其次，依据相邻尺度通道特征的可预测性，估算关键尺度的相邻尺度的通道特征，快速、高效地构建了多尺度聚合特征金字塔。然后，在Bootstrapping框架下，采用AdaBoost算法训练二阶决策树，构成行人分类器。最后，在进行行人检测时，按照预定的步长滑窗遍历每个尺度上的聚合通道特征，获得检测块，并将检测块作为训练好的级联分类器的输入，记录候选窗的窗口坐标及得分，利用非极大抑制对行人候选窗进行二次筛选，输出最后的行人检测框。在ETH和TUD等公开数据集进行测试，并与HOG方法、VJ方法、DPM方法相比较。研究验结果表明：DPM（形变部件模型）方法和提出的方法检测准确性高于VJ方法和HOG方法；在视角变化、行人存在遮挡的情况下，该方法在漏检、误检和窗口定位精度等方面的性能优于DPM方法，在保证较高检测精度的同时，极大地提高了检测速度，帧速率达到了29帧/s，优于其他算法，能够满足实时检测要求。<br>Amid at improving accuracy and speed of pedestrian detection， in the presence of background changes, measure uncertainty in pedestrian scale and occlusion was measured，based on the characteristics of the channel feature can be predicted reliably across adjacent scales， Therefore, a fast pedestrian detection method based on feature pyramid was proposed. First, channel features were aggregated at every key scale, which was composed of three LUV color space channels，one normalized gradient magnitude and six gradient direction histograms were calculated， which can fully reflected gradient and color information of the image. Second, based on the predictability of channel characteristics between adjacent scales， multi？？scale features of adjacent scales were estimated, and the feature pyramid was constructed quickly and efficiently. Then, with the bootstrapping framework, the AdaBoost algorithm was used to train second？？order decision？？trees to form a pedestrian classifier. Finally, during pedestrian detection, the aggregated channel features on every scale were segmented to obtain blocks, and the blocks were input to the trained cascade classifier. Window coordinates and scores of candidate windows were recorded. Non？？maximum suppression was used to screen the candidate windows. The final pedestrian detection box was output at last. The proposed algorithm was carried out on the ETH and TUD public datasets, and the results were compared with the HOG, VJ, and DPM methods. The results show that the proposed method and the DPM method have higher detection accuracies than the VJ and HOG methods. The performance of the method proposed in this paper is superior to the DPM method, regarding missed detections, false detections, and window positioning accuracy when the angle of view changes and in the presence of pedestrian occlusion. At the same time, 〖JP2〗the frame rate of the proposed method reaches 29 frames per second, which greatly improves detection speed and is able to meet real？？time
%K 交通信息与控制工程
%K 行人检测
%K 金字塔
%K 通道特征
%K 决策树&lt
%K br&gt
%K traffic information and control engineering
%K pedestrian detection
%K feature pyramid
%K channel feature
%K decision tree
%U http://zzszrb.chd.edu.cn/oa/DArticle.aspx?type=view&id=1805029