基于多特征融合的行人检测

研究了3 种不同类型的特征算子：梯度直方图（HOG），基于Gabor 变换的局部二值特征直方图（LGBPHS）和基于剪切波变换的直方图（HSC）在基于图像的行人检测中的应用。提出了基于多特征融合的检测算子，对单一特征进行L1 范式规格化之后，将3 个特征融合为一个高维的拥有大量信息的新特征，之后引入偏最小二乘法（PLS）进行特征降维，得到最终的人体特征。利用线性SVM 作为分类器，在INRIA 人体库上进行了实验，结果表明，融合后的特征极大的提高了检测率，在FPPW=10-5 时，检测率达到了95.6%。

Human Detection based on Multi Features Fusion

Based on the study of the applications of three different types of feature operatorsin human detection, which are Histogram of Oriented Gradient (HOG), Local Gabor BinaryPattern Histogram Sequence (LGBPHS) and Histogram of Shearlet Coefficients (HSC), wecombine them together and propose a new human detection feature operator. We employ PartialLeast Squares (PLS) analysis, an efficient dimensionality reduction technique, to project thefeature onto a much lower dimensional subspace. Using a linear SVM as the classifier, wecompare the fusion feature with the three single features in INRIA person dataset. Experimentsresults shows we achieve a detection rate of 95.6% with FPPW=10-5.