基于多特征融合的红外图像行人检测  

Pedestrian detection in infrared images based on multi-feature fusion

作　　者：胡庆新[1] 吕鹏[1] 

机构地区：[1]合肥工业大学计算机与信息学院,合肥230009

出　　处：《计算机应用》2016年第A01期157-160,195,共5页journal of Computer Applications

基　　金：教育部高等学校博士点基金资助项目(20120111110001);国家电子信息产业发展基金资助项目(工信部财函[2011]506)

摘　　要：针对基于单特征红外图像行人识别准确率低的问题,提出一种基于梯度方向直方图(HOG)、积分通道特征(ICF)和强度自适应特征(ISS)的多特征融合红外图像行人检测的新方法。首先,分别提取训练样本的HOG、积分通道和ISS特征,用主成分分析(PCA)算法对提取的ISS特征进行降维,然后通过并行加权特征融合方法把HOG、积分通道和降维后的ISS特征相融合,并用融合后的特征训练支持向量机(SVM)分类器,最后用训练好的SVM分类器进行行人识别检测。LSI Far Infrared Pedestrian Dataset红外行人图像数据库上的实验证明,基于多特征的红外图像行人检测方法明显优于经典的HOG和局部二值模式(LBP)单特征方法,提高了检测精度,降低了误检率。Focusing on the low accuracy of pedestrian detection in infrared images based on single feature,a pedestrian detection in infrared images method based on multiple features fusion that consists of Histogram of Oriented Gradient（ HOG）,Integral Channel Feature（ IDF） and Intensity of the Self-Similarity（ ISS） was proposed.First of all,the HOG,ISS and integral channel features of the training samples were extracted respectively,and due to the ISS feature＇ s dimensions is high,in order to increase the real-time performance,PCA algorithm was used to reduce the extracted ISS feature＇ dimension,and then the multi-feature fusion combination of HOG,integral channel feature and ISS feature was used after dimension reduction by the fusion method based on parallel weighting to train Support Vector Machine（ SVM） classifier,and the trained vector machine classifier was used to detect the Pedestrians finally.In the LSI Far Infrared Pedestrian Dataset database of Infrared pedestrian images,experimental results show that the pedestrian detection method in infrared images based on multi-features is superior to that based on the classic single feature method based on HOG and LBP,the detection accuracy is improved,and the false detection rate is reduced.

关 键 词：红外行人检测 特征融合 梯度方向直方图 积分通道特征 强度自相似特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]