欢迎访问《图学学报》
分享到：
摘 要：行人检测系统涉及交通安全问题，需要很高的鲁棒性，基于单特征结合单核支持向量机的方法效果有限，为解决这一问题，提出采用多特征和多核学习的方法来提升系统的鲁棒性，通过将积分信道特征、多层次导向边缘能量特征和CENTRIST特征分别与直方图交叉核、高斯核和多项式核进行线性组合，采用简单多核学习（Simple MKL）来分别计算核函数的权重系数，将多核学习方法与经典的梯度直方图特征/支持向量机、多尺度梯度直方图特征/直方图交叉核支持向量机和特征融合/直方图交叉核支持向量机的行人检测方法进行比较，实验表明所提出的行人检测算法的鲁棒性有明显提升。
关 键 词：简单多核学习；直方图交叉核支持向量机；CENTRIST特征；积分通道特征；多层
Abstract: Pedestrian detection system is involved in the traffic safety problem, and it requires very high robustness. The effect of the method based on single feature combined with single kernel support vector machine is limited. To solve this problem, this paper uses multi feature and multi kernel learning method to improve the robustness of the system. Integral channel features, multi-level oriented edge energy feature and CENTRIST features are respectively combined with histogram intersection kernel, gauss kernel and polynomial kernel. Simple multi-kernel learning (Simple MKL) is adopted to calculate the weight coefficients of kernel function. Multi-kernel learning pedestrian detection method is compared with histograms of oriented gradient feature/support vector machine, Multi-scale histograms of oriented gradient feature/histogram intersection kernel support vector machine and feature fusion/histogram intersection kernel support vector machine pedestrian detection methods. The experiments show that the robustness of pedestrian detection algorithm has obvious improvement.#br# Key words: simple multi-kernel learning; histogram intersection kernel support vector machine; CENTRIST feature; integral channel features; multi-level oriented edge energy feature
