基于多特征的AdaBoost行人检测算法  

AdaBoost for Pedestrian Detection Based on Multi-feature

作　　者：黄如锦[1] 李谊[1] 李文辉[1] 江琦[1] 杨赢涛[1] 

机构地区：[1]吉林大学计算机科学与技术学院,长春130012

出　　处：《吉林大学学报（理学版）》2010年第3期449-455,共7页Journal of Jilin University:Science Edition

基　　金：国家自然科学基金(批准号:60873147;60573182);国家高技术研究发展计划863项目基金(批准号:2008AA10Z224);吉林省科技发展计划项目基金(批准号:20060527)

摘　　要：针对单特征辨识度较低问题,基于多特征的AdaBoost行人检测算法,提出一种融合灰度和轮廓信息的新的多特征综合表示方法.该方法通过统计样本的权重直方图建立分类模型,并用多个直方图的乘积表示样本在多特征下对应的联合概率分布,从而基于多特征联合概率更精准地描述行人,提高行人检测的鲁棒性.实验结果表明,改进后的基于多特征行人检测算法提高了行人检测精度、降低了误检率,目标识别的置信度明显提高,在多变的自然背景下可以取得较好的效果.In view of the problem of the low degree of single-feature recognizability in pedestrian detection based on feature,this paper introduces a kind of algorithm,AdaBoost for pedestrian detection based on multi-feature,and presents a new integrated approach of multi-feature which is the fusion of gray scale and contour information. This approach establishes a classification model by the histogram statistics of the weight samples,and the probability distribution is represented by multiplication of several histograms. So the joint probability based on multi-feature can get more accurate description of pedestrian,and improve the robustness of pedestrian detection. The experimental results show that pedestrian detection in this paper has improved the detection rate,lowered false alarm rate in a large extent,and the confidence level of target recognition has been markedly improved,and it can get the better performance of pedestrian detection in various natural background.

关 键 词：行人检测 多特征 直方图统计 查找表 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]