摘要
外观、尺度变化是行人跟踪的难点,解决行人多尺度跟踪问题是增强算法实用性的关键因素.在KCF(kernel correlation filter)算法的基础上,本文采用多个相关滤波器(如头部、臀部)辅助身体躯干滤波器的匹配跟踪.通过获得图像帧(除第一帧外)与初始帧的行人头部和臀部之间的距离变化率来缩放搜索面积,解决目标定位不准确和时间浪费的问题;通过调整目标框的尺寸,解决目标模板逐渐包括背景特征或者逐渐被局部特征取代的问题.在VOT2016的18个有明显尺度变化的行人场景视频序列上进行了测试,实验结果表明所提算法具有更高的跟踪准确率.
Appearance and scale change are the difficulties of pedestrian tracking. To solve the problem of multiscale pedestrian tracking is the key factor to enhance the practicability of the algorithm. On the basis of KCF(kernel correlation filter) algorithm,this paper uses multiple correlation filters(such as head and hip)to assist the tracking of the body trunk filter. The distance change rate which is obtained by comparing the distance between the pedestrian’s head and hip of every image frame(except the first frame)with the initial frame is used to zoom the search area, so as to avoid inaccurate target location and time waste. By adjusting the size of the target’s bounding box, the problem of target’s template shift caused by gradual change of the target template including background features or local features is solved. The experiment is conducted on eighteen pedestrian scene video sequences with obvious scale changes in VOT2016 dataset, and the experimental results show that the algorithm proposed has higher tracking accuracy.
作者
张云洲
郑瑞
暴吉宁
朱尚栋
ZHANG Yun-zhou;ZHENG Rui;BAO Ji-ning;ZHU Shang-dong(School of Information Science & Engineering,Northeastern University,Shenyang 110819,China)
出处
《东北大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2019年第9期1228-1233,1239,共7页
Journal of Northeastern University(Natural Science)
基金
沈阳市高层次创新人才支持计划项目(RC170490)
中央高校基本科研业务费专项资金资助项目(N172608005,N182608004)
国家自然科学基金资助项目(61471110,61733003)
