Abstract: This paper presents a feature fusion method based on Local Difference Binary Pattern（LDBP） and Local Binary Pattern（LBP） for solving the problem of the inaccurate and unstable pedestrian detection. Firstly, it filters the input image by using the 2D discrete haar wavelet transformation, in order to obtain four sub-images with different frequency, including LL, LH, HL and HH. Secondly, it extracts the LDBP feature from the low frequency part and the LBP feature from the other three high frequency parts. Then, it reduces the dimension of the LDBP and LBP feature spaces by using Principal Component Analysis（PCA）. Finally, it applies the fused LDBP-LBP feature for performing an efficient pedestrian detection. Conducted on INRIA databases by using Support Vector Machine（SVM）, the experimental results demonstrate this method can effectively improve the detection accuracy and get better robustness.

Key words: pedestrian detection, Local Difference Binary Pattern（LDBP） feature, Local Binary Pattern（LBP） feature, feature fusion

摘要： 提出一种基于局部差分二值模型（Local Difference Binary Pattern，LDBP）和局部二值模型（Local Binary Pattern，LBP）的特征融合方法，以解决行人检测中检测精确度和鲁棒性不足的问题。对输入图像进行二维离散Haar小波变换，得到不同频率的四个子图像（LL，LH，HL和HH）；对低频部分子图像提取LDBP特征，以及对其他三个高频部分子图像提取LBP特征；采用主成分分析法（PCA）分别对得到的LDBP特征和LBP特征进行降维；融合降维后的LDBP特征和LBP特征进行行人检测。在INRIA数据集上采用支持向量机（SVM）进行测试，实验结果表明，该方法能有效地提高检测精确度，且具有较好的鲁棒性。

关键词: 行人检测, 局部差分二值模型（LDBP）特征, 局部二值模型（LBP）特征, 特征融合