视角和旋转角变化时梯度方向直方图的转换

The conversion of histograms of oriented gradient in different vision-angle and rotation-angle

使用梯度方向直方图(HOG)来检测目标,需要大量的,有代表性的样本来训练分类器.一个目标的HOG,其特征在不同的摄像机视角和不同的光轴旋转角下,并不相同.因此,使用不同视角下的混合样本集来训练分类器时,目标检测的准确率受到样本噪声的影响将会降低.基于摄像机成像的基本原理,提出了一种转换算法,可以把一个样本在某个视角下的HOG特征转换成另一个视角下的HOG特征.这样既降低了分类器训练时需要采集的正负样本数量,又提高了支持向量机(SVM)分类的准确性,从而提高了目标检测的准确性.大量目标检测实验结果表明本文提出的算法是有效的.

In applying the histograms of oriented gradient（HOG） to detect an object, we need a great number of representative image samples to train the classifier. Since the HOG characteristic changes in different vision-angle and different rotation-angle, the detection accuracy will be decreased if images of different vision-angle or rotation-angle are used to train the classifier. By the imaging principle of the camera, we develop an algorithm for converting the HOG characteristic in one vision-angle and rotation-angle to the HOG characteristic in another vision-angle and rotation-angle. Thus, the required number of positive and negative samples for training the classifier is reduced and the classification accuracy of the support-vector-machines（SVM） is raised, eventually resulting in an increase in the object detection accuracy and robustness. Many object-detection experimental results show that this conversion algorithm is effective. This indicates that the proposed algorithm is an efficient tool for HOG-based object detection in practical engineering projects.