摘要
针对多变化场景中通用分类器复杂度高和手工标记工作量大的问题,提出了一种新的迁移学习框架,结合稀疏编码和背景差分进行行人检测。首先优化HOG+SVM通用检测器,融合BCLBP和HOG进行特征提取,训练linearSVM,并在目标场景序列上利用基于GMM的背景差分法获得帧目标样本的运动区域以丰富样本特征。其次利用尺寸等过滤器从目标样本中筛选出部分样本作为目标模板,然后通过稀疏编码计算源样本与目标样本和目标模板的相关性,根据稀疏系数与置信度值去计算源样本和目标样本的权重。在重训练过程中,基于稀疏编码对所有样本进行权重分配,排除源样本的异常点,从而解决目标样本漂移,得到特定场景的行人检测器。为验证算法的有效性,在INRIA、Caltech、TUD数据集上实验,本文训练的特定场景行人检测器的检测率相对于其他传统方法实现了不同程度的提高。
Aiming at the problem of high complexity of generic detector and large workload of manual labeling,a new transfer learning framework is proposed,which combines sparse c oding and background subtraction to detect pedestrians.Firstly,the HOG＋SVM generic detector is o ptimized,BCLBP and HOG are merged to extract the features and train linear SVM.The GMM-based ba ckground subtraction method is used to obtain the motion region of the frame target samp les to enrich the sample features.Secondly,use the filters including size to filter out some s amples from the target sample as the target template,and then calculate the correlation between the source sample and the target sample and the target template by sparse coding.According to the sparse coefficient and confidence value,calculate the weight of source samples and target samples.I n the process of retraining,all the samples are with allocated weight based on sparse coding,and the abnormal samples of the source samples are excluded,so as to solve the target sample drift and obtain t he pedestrian detector of a specific scene.In order to verify the validity of the algorithm,experiments on the INRIA,Caltech and TUD datasets show that the detection rate of the pedestrian detector in our spec ific scene is improved to some extent compared with other traditional methods.
作者
崔鹏
赵莎莎
CUI Peng;ZHAO Sha-sha(School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China)
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2018年第9期1012-1020,共9页
Journal of Optoelectronics·Laser
基金
黑龙江省自然科学基金(F2015038)
黑龙江教育厅(11551086)资助项目
关键词
迁移学习
稀疏编码
背景差分
行人检测
,,transfer learning
sparse coding
background subtraction
pedestrian detection
