行人检测中非极大值抑制算法的改进

Improvement of Non-maximum Suppression in Pedestrian Detection

行人检测是计算机视觉领域的难点和热点问题。行人检测可大致划分为3个部分:特征提取、分类和非极大值抑制(Non-maximum Suppression,NMS)。当前的研究工作主要集中在特征提取、特征学习和分类器等方向,而非极大值抑制方向鲜有改进。目前常用的非极大值抑制算法是贪心策略,抑制时只使用了单一的重合面积信息。针对该问题,在ACF(Aggregate Channel Features)检测算法的基础上,对非极大值抑制进行了3项改进,显著地提升了算法的精度,并且算法的时间消耗只有略微的增加。在INRIA数据集上,单独使用引入尺度比的动态面积阈值NMS时能降低平均对数漏检率(MR)0.99%;单独使用保留外围检测分数相近的检测窗口的策略时NMS能降低MR 1.25%;两者结合可降低MR 2.5%;结合后,再对已经被抑制的检测窗口重复抑制,MR降低了2.63%,达到14.22%。

Pedestrian detection is a hot topic in computer vision and it includes three parts.- feature extraction, classification, and non-maximum suppression （NMS）. Most of the existing results on pedestrian detection have focuses on feature extraction, feature learning, and classifier. To the contrary, there exist few results on NMS. Moreover, the common approach on NMS is based on greedy strategy, which only uses the information of overlapping area to suppress other windows. By means of ACF （Integral Channel Features）, this paper proposes three improved NMS algorithms, which can notably raise the accuracy of computation without increasing computation time. On the data set of INRIA pedestrian, when only using the dynamic overlap threshold changing with the scale rate, the MR （log-average miss rate） can be reduce by 0.99%, which can be only reduced by 1.25% when using the strategy of saving outlying detection windows with the similar score. The integration of these two methods can reduce MR by 2.5 %. Furthermore, the suppression again of the suppressed detection windows can further reduce MR by 2.63% to achieve a lower MR with 14.22%.