基于非稀疏多核支持向量机的重现行人识别  

Pedestrian Re-identification Algorithm Based on Non-sparse Multiple Kernel Support Vector Machine

作　　者：许允喜[1,2] 齐赵毅[1] 

机构地区：[1]湖州师范学院信息与工程学院,浙江湖州313000 [2]浙江大学信息与电子工程系,杭州310027

出　　处：《光电工程》2013年第9期82-87,共6页Opto-Electronic Engineering

基　　金：国家自然科学基金项目(60872057);湖州市自然科学基金项目(2011YZ07);国家级大学生创新创业训练计划项目(201210347010);湖州师范学院校级科研项目成果(KX24061)

摘　　要：目前,视频跟踪正向大范围长时间目标跟踪研究方向发展。重现行人识别是对行人目标进行大范围长时间持续跟踪的关键技术,是后续行为分析的基础。本文提出了一种基于非稀疏多核支持向量机的重现行人识别算法。首先,该方法提取跟踪行人视频图像序列的多层SIFT视觉单词树特征和多层颜色直方图特征。接着,利用高效的非稀疏多核支持向量机算法在线融合多层SIFT视觉单词树特征和多层颜色直方图特征得到行人外观模型。最后利用存储的行人外观模型库对重现行人进行识别。该方法可应用于多摄像机视频监控中同一行人目标的跨摄像机跟踪以及单摄像机监控中行人目标重新出现的识别。实验结果表明,该方法能快速训练人体目标外观模型,能获得很高的识别率。The research of video tracking is developing forward wide-range and long-time object tracking. Pedestrian re-identification is the key technology of wide-range and long-time pedestrian tracking, and is foundation of follow-up behavior analysis. A pedestrian re-identification algorithm is proposed based on non-sparse multiple kernel Support Vector Machine （SVM）. Firstly, we extract multilayer SIFT feature and multilayer color histogram feature of tracked pedestrians video image sequence. Then, we online fuse multilayer SIFT feature and multilayer color histogram feature to obtain pedestrian appearance models using non-sparse multiple kernel SVM. Finally, we re-identify pedestrian objects using the stored pedestrian appearance models. The method can be applied to the same pedestrian tracking across cameras in the multiple cameras video surveillance and recognition of pedestrian recurrences in the single camera video surveillance. The experiment results show that our method can rapidly train pedestrian object appearance models and achieve very high recognition rate.

关 键 词：智能视频监控 多核支持向量机 单词树 行人重现识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]