【摘要】 针对现有行人检测算法未考虑正负样本非均衡性及分类器间所需的差异性的不足,提出一种同时考虑分类器多样性及正负样本非均衡性的行人检测算法.首先,在分类器中引入代价敏感的思想,通过设置适宜的代价敏感参数值,使分类器更加关注数量较少且更为重要的行人正样本;进而,通过动态调整对分类算法性能影响较大的核函数参数σ的选择,形成一组相互间有差异且分类性能适度精确的一组分类器,并对分类器之间的相似度进行对比,剔除相似度高的分类器;最后,将剩余分类器级联组合.实验证明,和经典算法相比,提出的算法有利于提高行人检测精度,且虚警率更低.更多还原

【Abstract】 A pedestrian detection algorithm take the non-equilibrium between positive and negative samples and the differences of classifiers into account at the same time is proposed,for which the existing algorithm that does not consider.Firstly,introduce the idea of cost-sensitive into classifier by setting cost-sensitive parameter′s value,and make the classifier pay more attention to positive samples which fewer and more important.Furthermore,dynamic adjustment the kernel function parameterσwhich greater impacts on the performance of the classification algorithm, and make up a set of classifiers which mutual difference and moderately accurate.Then exclude the high similarity classifier by contrast.Finally,cascading and composing the remaining classifiers.Experiments show that compared with the classical algorithm,the proposed algorithm is conducive to improve pedestrian detection accuracy and has lower false alarm rate.更多还原