视角变化情况下目标的HOG特征转换算法研究
首发时间:2008-11-14
摘要:从视频图像中检测出人、车辆等目标是计算机视觉应用的一个关键步骤,基于梯度方向直方图(HOG)的检测方法已经被证明具备足够的鲁棒性和良好的检测效果。由于HOG方法需要大量的、具备足够代表性的样本来训练分类器,而同一个目标的HOG特征在不同的摄像机视角、不同的旋转角下并不相同,使用不同视角下的混合样本集来训练分类器,目标检测的准确率受到样本噪声的影响而偏低。因此,在训练样本不足的情况下,通过某个确定视角下样本的HOG特征,推算出另一个视角所对应的HOG特征,成为HOG目标检测器在实际工程应用中需要解决的关键问题。本文提出了一种不同俯仰角、不同目标旋转角和不同光轴旋转角下,样本HOG特征的直接转换算法,以提高HOG检测的效果和鲁棒性,提高SVM分类效果,降低分类器训练时需要采集的正负样本的数量。实验结果表明本文提出的算法是正确和有效的。
关键词: 目标检测 视角 梯度方向直方图 SVM 计算机视觉
For information in English, please click here
The Transition Algorithm for Object’S HOG Under Visual Angle Change
Abstract:Finding object in video images is critical for several applications in computer vision, such as intelligent surveillance systems. A detection algorithm based on histograms of oriented gradient (HOG) has proved which has enough robustness and good results of detection. HOG detection method need vast and representational image sample to train classifier, but the same object’s HOG is different in different visual angle. So it is the urgent key problems that transform the sample’s HOG from the certain visual angle to another visual angle in application of HOG detection method in practical projects. This paper proposes a transition algorithm for the sample’s HOG in different pitching angle, rotation angle of object and rotation angle of optical. Experiments proved the new algorithm can improve the results of SVM classify and the accuracy of HOG algorithm, also reduced the quantity of positive and negative sample for training. These tests show that the method is accurate and effective.
Keywords: object detection different visual angle histograms of oriented gradient SVM computer vision
论文图表:
引用
No.2572728983012266****
同行评议
共计0人参与
勘误表
视角变化情况下目标的HOG特征转换算法研究
评论
全部评论0/1000