基于局部二值模式和四元数的行人检测

局部二值模式(LBP)特征具有光照不变性、旋转不变性及计算简单等特性,能有效表示行人特征,广泛应用于行人检测。LBP 特征的提取方法通常基于灰度图像,如果用于彩色图像,则没有充分考虑各通道之间的相关性,不能保证行人检测的准确性。为此,我们对彩色图像的红、绿、蓝三通道的LBP 特征,以四元数的形式表示,利用四元数的性质,提出一种P-LBP特征,再利用k-最近邻算法训练分类器。该方法在INRIA 数据集上进行实验,与HOG、S-LBP、F-LBP、HOG-LBP特征进行比较,具有更好的效果。

Pedestrian Detection Based On Local Binary Pattern And Quaternion

Local binary pattern (LBP) feature owns the properties of invariability in illumination and rotation, simplicity in calculation. Therefore it can describe pedestrian effectively, and is widely used in pedestrian detection. Most existing methods usually extract LBP features from gray images, if these methods are used in the color image, they can"t guarantee the accuracy of pedestrian detection for the reason of neglecting the correlation between each color channel. Therefore, we use the quaternion to express the LBP features of red, green and blue channels which are extracted from the color images, put forward a P-LBP feature by taking advantage of the quaternion"s property, and then use the k-neighbour algorithm to train the classifier. Comparing with the HOG, S-LBP, F-LBP and HOG-LBP features, the experiments in the INRIA Dataset show that our method perform better.