基于仿射梯度方向直方图特征的目标识别算法  

The Object Recognition Algorithm Based on Affine Histogram of Oriented Gradient

作　　者：宋丹[1] 唐林波[1] 赵保军[1] 

机构地区：[1]北京理工大学信息与电子学院,北京100081

出　　处：《电子与信息学报》2013年第6期1428-1434,共7页Journal of Electronics & Information Technology

摘　　要：针对基于传统梯度方向直方图特征的目标识别算法(HOG+SVM)在目标发生仿射变化时识别效果较差的问题,该文提出一种基于仿射梯度方向直方图特征的目标识别算法(AHOG+SVM)。通过提取多尺度金字塔梯度图像的HOG特征,提高了算法的尺度不变性;通过将平面HOG栅格拓展至3维HOG栅格,并根据目标的世界坐标系与图像坐标系的映射关系将3维HOG栅格映射为2维HOG仿射栅格,最后对仿射栅格内的HOG特征进行仿射逆变换,以达到增强算法旋转不变性与错切不变性的目的。多组实验结果表明,该文提出的算法能够解决在目标识别过程中由尺度变化、旋转变化和错切变化(3D视角变化)所造成的识别率较低的问题,性能优于HOG+SVM算法。A kind of object recognition algorithm based on Affine Histogram of Oriented Gradient （AHOG＋SVM） is proposed to solve the poor effect of object recognition algorithm based on HOG （HOG＋SVM）. In order to have scale invariance, this paper builds nmlti-scale pyramid gradient images, and then computes HOG feature on them. In order to increase the rotational invariance and shear invariance, this method firstly expands 2D HOG grid to 3D HOG grid, then maps 3D grid to 2D affine grid according to the relationship between the world coordinate and image coordinate. Finally, inverse transformation of HOG feature in affine grid is carried out to remove the influence of affine transformation. The experimental results show that, the proposed method has the ability to solve the low recognition rate because of scale changes, rotation changes and shear changes （3D perspective changes） of object, and its performance is better than HOG＋SVM.

关 键 词：目标识别 仿射变换 梯度方向直方图特征 仿射栅格 视角变化 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]