HOG结合傅立叶局部二值模式特征的行人检测

Pedestrian detection algorithm combining HOG with LBPHF

针对局部二值模式(Local binary patterns,LBP)和梯度方向直方图(Histogram of Gradients,HOG)的联合特征,在行人检测中易受行人肢体偏转影响的问题,本文将傅立叶局部二值模式算子(Local binary patterns-HF,LBPHF)与HOG算子联合对行人进行特征描述。在每个滑动窗口中,分别计算HOG特征与LBPHF特征,将两者结合,构成联合特征。利用线性支持向量机训练分类器,通过自举法不断更新优化分类器,获得最优判别模型。将提取所得的联合特征输入分类器中进行判别,采用非极大值抑制的融合方法对重叠检测窗口进行融合。实验结果表明,LBPHF算子与HOG相结合的方法检出率高,计算复杂度低,抗行人肢体偏转干扰能力强。

The augmented feature set combining the fundamental local binary pattern (LBP) with the histograms of oriented gradients (HOG) would be easily affected by limb deflection of pedestrian in pedestrian detection. A novel augmented feature set is proposed combining the local binary pattern histogram based on Fourier transform (LBPHF) with HOG. The complete pedestrian detection system can be divided into two phases. HOG and LBPHF could be calculated respectively in each sliding window to construct the feature set. Subsequently, the linear support vector machine (SVM) is adopted to train the classifier. In order to obtain the optimal discriminant model, Bootstrap is used to update the classifier. In detecting phase, the augmented feature set will be put into the classifier. Then the Non-maximum suppression is adopted for over lapping detections. The experiment results on the INRIA personal dataset show that the proposed algorithm can achieve higher detection rate, lower computational complexity and a strong anti-interference performance of limb deflection.