从正面车辆图像中识别车型的中层特征表示方法

为有效识别正面车辆图像中的车辆类型,提出了一种车辆中层特征表示方法. 该方法以SIFT特征为底层特征,建立两类中层特征:结构特征分布和表观特征分布. 结构特征分布是对SIFT特征在图像中位置信息的统计,描述车辆不同部分之间的相对位置关系,在不同类车辆之间以及车辆与背景之间具有较强的判别力;表观特征分布是对SIFT特征本身的统计,描述车辆的局部表观信息,对光照变化和背景干扰十分鲁棒. 利用多核学习方法融合两类中层特征,进而识别车辆类型. 实验结果表明,该方法在光照变化和有背景干扰等复杂情况下表现出良好的车型识别性能. 

Middle-Level Features for Vehicle Classification from Frontal View Images

Based on the SIFT feature, two kinds of middle-level features were proposed for vehicle type classification from vehicle's frontal view images. Structural feature distribution characterizing the spatial relative layout of vehicle parts helps to distinguish different types of vehicles from the background, and the appearance-based feature distribution is local and robust to the interference of illumination variation and the background. These two kinds of distributions were combined to classify vehicle types by multiple kernel learning. The experimental results demonstrate that our method is robust to various illumination and background interference.