视频监控中基于在线多核学习的目标再现识别

在非重叠多摄像机或单摄像机视频监控中,识别跟踪目标的再次出现很重要.针对传统支持向量机方法在特征融合方面的缺陷,本文提出了一种新的基于在线多核学习的人体目标再现识别方法.该方法对跟踪目标视频前景图像序列提取具有互补性的视觉单词树直方图和全局颜色直方图二种特征,再采用多核学习方法在线训练人体目标视觉外观,从而得到多核特征融合模型.实验结果表明,该方法能快速训练人体目标外观模型,满足视频监控的实时要求,多核融合模型获得了比单一特征模型和单核支持向量机方法更高的识别性能.

People Re-identification Based on Online Multiple Kernel Learning in Video Surveillance

In the non-overlapping multi-camera or single camera video surveillance, re-identification of tracked target is very important. Due to weakness of traditional support vector machine in feature fusion, a new people re-identification method is proposed based on online multiple kernel learning. We extract complementary visual word tree histogram and global color histogram from tracked people foreground image sequence in video, and then multiple kernel learning method is used for online train people visual appearance. Finally, we obtain multiple kernel feature fusion model of people appearance. Experimental results show that our method can train people appearance model rapidly, meet the real-time requirement of video surveillance, and attain higher recognition performance than single feature appearance model and single kernel support vector machine method.