航拍图像车辆检测中的圆形滤波器HOG特征快速计算

航拍图像中车辆一般近似为矩形结构，因此通过统计检测窗口中的梯度方向直方图并根据梯度主方向估计车辆朝向，将检测窗口旋转到相应方向进行分类器判别。车辆检测采用级联boosting分类器和梯度方向直方图特征，针对旋转窗口中梯度方向直方图特征的计算，设计一种基于圆形滤波器的梯度方向直方图特征。与传统基于积分直方图的梯度方向直方图特征提取方法相比，显著提高了旋转窗口中梯度方向直方图特征的计算效率，在计算每个像素的梯度时采用查找表代替梯度向量的求模和角度计算也减小了计算量。使用真实图像进行的实验表明，该车辆检测算法快速高效。
In general, cars are rectangular shape in the aerial images, so the histograms of orient gradient over the whole sliding window were computed to find the primary gradient direction and to estimate the orientation of the car in the window, and the detection window was rotated according to the car’s orientation to perform classification. A cascaded boosting classifier and the HOG (histograms of orient gradient) features in the proposed car detection method were employed. To efficiently compute the HOG features in the rotated window, a fast HOG features extraction method based on CFHOG (circle filter based histograms of orient gradient), which was more efficient than the classical HOG extraction method based on integral histograms. In addition, look up tables are used to speed up the calculation of the orientation partition and magnitude. A set of experiments on real images prove the applicability and high efficiency of the proposed car detection method.