多卷积特征融合的HOG行人检测算法  

HOG Pedestrian Detection Algorithm of Multiple Convolution Feature Fusion

作　　者：高琦煜 方虎生 

机构地区：[1]陆军工程大学,南京210007

出　　处：《计算机科学》2017年第B11期199-201,232,共4页Computer Science

基　　金：国家973计划资助项目(2009CB723803);国家自然科学基金资助项目(60873120)资助

摘　　要：行人检测是计算机视觉领域中的经典问题,HOG结合SVM的方法是解决这一问题的有效途径,HOG对行人特征的有效描述起到了重要作用。卷积神经网络(CNN)作为一种有效的特征提取方法,通过特征图可以实现对特征更好的描述。提出将卷积神经网络(CNN)与传统的HOG+SVM算法相结合的方法。首先利用CNN在下采样层中可以使用不同的卷积核对数据进行不同角度特征描述的特点,对样本进行多角度浅层特征提取;然后用HOG对得到的浅层特征进行进一步的提取;最后采用支持向量机(SVM)完成训练、分类。实验表明,该方法对于行人检测具有很高的识别率,优于传统方法。Pedestrian detection is utilized as the fundamental of various computer vision applications. A typical and ef-fective solution of pedestrian detection is combining histogram of oriented gradient （HOG） with support vector machine （SVM）. In this paper, we proposed a novel pedestrian detection method, which uses convolutional neural network （CNN） in HOG＋SVM to obtain more comprehensively feature description through a variety of convolutional kernels in CNN. Firstly,it extracts shallow features from data set by CNN, and uses CNN^s characteristics of displacement, and scale and deformation invariance. Then, it merges strongly relevant features by analyzing correlation coefficient of shal-low features. After this, it extracts HOG features from CNN shallow features which are weakly correlated with each other. Finally, it uses SVM to complete the training and classification. The experimental results show that the proposed method can obtain higher accuracy than the existing method in pedestrian detection.

关 键 词：行人检测 卷积神经网络 多卷积特征 HOG 

分 类 号：TP139.41[自动化与计算机技术—控制理论与控制工程]