’ Histograms of Oriented Gradients for Human Detection ’ versus ’ Fast Human Detection Using a Cascade of Histograms of Oriented Gradients
’ Histograms of Oriented Gradients for Human Detection ’ versus ’ Fast Human Detection Using a Cascade of Histograms of Oriented Gradients
复制标题
DOI：
--
复制
发表时间：
2007
期刊：
影响因子：
--
通讯作者：
Qiang Zhu;Mei-Chen Yeh;K. Cheng;S. Avidan
中科院分区：
文献类型：
--
作者：
Qiang Zhu;Mei-Chen Yeh;K. Cheng;S. Avidan
Dalal和Triggs等研究了鲁棒视觉对象识别的特征集问题。他们首先考虑了现有的基于边缘和梯度的描述符，然后他们通过实验证明了定向梯度直方图（HoG）描述符的网格明显优于现有的人体检测特征集。在此之后，他们研究了计算的每个阶段对性能的影响。他们最终得出结论，在重叠描述子块中，精细的尺度梯度、精细的方向分束、相对粗糙的空间分束和高质量的局部对比度归一化是获得良好结果的重要因素。他们的探测器在原始的麻省理工学院行人数据库中获得了近乎完美的结果，因此他们制作了一个更具挑战性的数据集，名为INRIA，其中包含极其复杂的背景和戏剧性的照明变化。朱强，Shai Avidan， Mei-Chen Yeh和Kwang-Ting Cheng等人随后表明，将级联拒绝器方法与HoG特征相结合，可以实现快速准确的人体检测系统。特征是具有可变块大小的hog。为了从更大的可能块集中选择最佳块进行检测，使用AdaBoost （Adaptive Boosting的缩写）。此外，采用积分图像表示，抑制级联显著提高了计算速度。
Dalal and Triggs [1] studied the question of feature sets for robust visual object recognition. They first considered existing edge and gradient based descriptors and then they showed experimentally that grids of Histograms of Oriented Gradients (HoG) descriptors significantly outperform existing feature sets for human detection. After this they studied the influence of each stage of the computation with reference to performance. They finally came to the conclusion that fine scale gradients, fine orientation binning, relatively coarse spatial binning, and high quality local contrast normalization in overlapping descriptor blocks were important to get good results. Their detector reached nearly perfect results on the original MIT pedestrian database, so they produced a more challenging dataset called INRIA which contained extremely complicated backgrounds and dramatic illumination changes. Qiang Zhu, Shai Avidan, Mei-Chen Yeh, and Kwang-Ting Cheng [2] then showed that the combination of the cascade of rejectors approach and the HoG features led to a fast and accurate human detection system. The features were HoGs with variable block-size. To select the best blocks for the detection, from a larger set of possible blocks, AdaBoost (short for Adaptive Boosting) was used. Furthermore, the integral image representation was used and the rejection cascade significantly speeded up the computation.