摘要
针对AdaBoost(Adaptive Boosting)算法耗时长,检测精度低的问题,提出一种将支持向量机(Support Vector Machine,SVM)与AdaBoost算法相结合的方法,采用多尺度(Histogram of Oriented Gradient,HOG)特征描述行人不同尺寸区域的特点,在样本权重更新过程中引入非行人样本的误检率,训练了一个级联结构的行人分类器.结果表明:选择特征数目为5个时,改进后的AdaBoost算法检测率和计算时间分别为99.8%和577.66s.在INRIA行人数据库上训练了一个六层结构级联分类器,使用特征数目为46个时,检测率达到96.74%,而误检率仅有3.26‰.
For the long training time of AdaBoost and low detection accuracy of algorithm,this paper proposes a combined algorithm based on Support Vector Machine and AdaBoost.Multi-size HOG features is adopted to describe different size areas of the characteristics of person,false inspection rate of non-person sample is introduced in the process of sample weight update,a cascade structure of the pedestrian classifier is trained.Under the condition of 5 selected features,detection accuracy and calculation time of modified AdaBoost algorithm are 99.8% and 577.66s separately.A six-level cascade classifier is trained,using a total of 46 features,and it achieves a detection accuracy of 96.74% while the false inspection rate is only 3.26 ‰ on INRIA person database.
出处
《西安工业大学学报》
CAS
2012年第4期259-263,共5页
Journal of Xi’an Technological University
基金
西安市科技局项目(CXY1015-1)
陕西省教育厅专项科研计划项目(11JK0991)