【摘要】 为了提高车辆视频检测和训练速度,针对车辆视觉特征维数较高的普遍问题,构造了一种特征降维方法,采用偏最小二乘(PLS)法对包含正、负样本的训练集图像进行分解;通过变量投影重要性分析(VIP)法评价原始特征对分解结果的贡献得分,将得分降序排序并选用高分特征直接张成低维空间,实现数据降维,并以低维数据为输入项,学习得到车辆分类器。对于新图像,直接抽取对应的高分特征进行检测,避免了常见的数据投影过程。研究结果表明:采用PLS-VIP降维,车辆图像的聚类性更明显,聚类正确率优于传统的主成分分析法、PLS法、非线性降维的等距映射法、特征脸法以及PLS衍生方法,降维后检测耗时达到降维前的22%。更多还原

【Abstract】 In order to improve the speed of training and visual detection for vehicles,aimed at the extremely high dimensional features,a novel dimensionality reduction method was proposed.The partial least square(PLS)method was introduced to decompose the training images of both negative and positive samples,the VIP was applied to estimate the scores of features in decomposition,and the dimensionality reduction was finished when scores were sorted in descend order to select the leading features to span the low-dimensional space.Data in low-dimensional space were taken as the input for the learned classifier of vehicles.As to the new images,due to the direct use of leading features guided by scores,the common data projections were no longer involved.The results show that the PLS-VIP based dimensionality reduction method is more suitable for clustering and is able to obtain higher clustering accuracy on vehicle images than traditional methods,such as PCA,PLS,ISOMAP,eigenface,and PLS-based methods.The detection time is reduced to 22% of that before dimensionality reduction.更多还原