基于多模型检测的视频监控行人跟踪算法  

Pedestrian tracking algorithm by video surveillance based on ensemble detection

作　　者：韩光星 张轩[1] 李崇荣[1] Han Guangxing Zhang Xuan Li Chongrong(Institute for Network Sciences and Cyberspace, Tsinghua University, Beijing 100084, China)

机构地区：[1]清华大学网络科学与网络空间研究院,北京100084

出　　处：《华中科技大学学报（自然科学版）》2016年第11期43-47,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(60703052)

摘　　要：研究了网络视频监控中行人检测和跟踪技术,并提出了基于多模型检测的视频监控行人跟踪算法.首先在已有检测算法的基础上,根据实验结果,分析了算法的优势和存在问题,提出了多模型融合检测算法,多模型融合检测算法充分利用各个单模型算法的优势,提高模型检测准确率,并使用快速特征金字塔算法提高了算法的实时性.随后在提出的检测算法基础上,结合卡尔曼滤波算法和匈牙利最优匹配算法实现多目标行人检测和实时跟踪.MOT测试视频实验结果表明:提出的跟踪算法能够较好地实现多目标行人检测和跟踪,适合网络视频监控场景.The pedestrian detection and tracking problem in network video surveillance were studied,and a pedestrian tracking method was presented based on ensemble detection algorithm.Firstly,the strengths and weaknesses of state-of-the-art detection algorithms were evaluated according to experimental results.Meanwhile,an ensemble model making full use of the existing algorithms was proposed which greatly improved the detection accuracy.By using the fast feature pyramid method,the real-time performance of the proposed detection algorithm was further improved.Then,by combining the Kalman filter and Hungarian method,the pedestrian tracking algorithm was proposed.Experiments on MOT challenge video data demonstrate that the proposed method can robustly detect and track multiple pedestrians simultaneously,which is suitable for video surveillance scene.

关 键 词：网络视频监控 行人检测 行人跟踪 卡尔曼滤波 匈牙利算法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]