多传感器人体检测的FHOG图像特征融合

提出了一种新的基于方向梯度直方图(HOG)的图像特征融合方法. 该方法采用视觉激活度(VAM)来选择具有显著方向性的局部梯度统计值,构成融合的方向梯度直方图(FHOG),有效地解决了多分辨率(MR)图像融合存在的不足. 文中把这些融合特征输入线性支持向量机(SVM),训练得到人体/背景二元分类器用于人体检测. 实验表明,与传统多分辨率图像融合方法相比,在参考点处本文提出方法漏检率下降3～10%,虚警率平均下降20%以上. 

Image Feature Fusion for Human Detection with Multi-Sensor Based on FHOG

Proposed a novel image feature fusion approach based on histogram of oriented gradient (HOG). The visual activation measure (VAM) was used to select the statistics of local gradients with significant direction, and forms fused histogram of gradient (FHOG), which effectively solved the existing deficiencies of multi-resolution (MR) image fusion. These fused features were plugged into support vector machine (SVM), and train human/background binary classifier for human detection. Experiments show that compared with the traditional MR image fusion approaches, in reference points the missing rate of the proposed approach decreases by 3%～10%, and the false alarm rate drops by an average of more than 20%.