摘要
基于SiamFC跟踪网络和Faster R-CNN检测网络,提出基于关键帧模板更新算法,提高行人目标跟踪速度,以及根据跟踪目标连续性和目标形状不会突变性质,提出欧氏距离和重叠度对系统跟踪结果进行约束。经实验验证,算法系统跟踪平均重叠度达到0.82,算法中心点l2-norm距离缩小到9,有效提高系统的跟踪质量,优于基于检测或者跟踪算法,在系统配置为NVIDIA 1080Ti显卡下,系统跟踪速度达到37fps,达到实时跟踪效果。
Based on SiamFC Network and Faster R-CNN Network,proposes an algorithm of updating the template of the key frame to improve the speed of pedestrian tracking.Since the pedestrians are continuous and invariant,uses l2-norm distance and the IOU of the detecting or track⁃ing target box to constraint the result.The progressiveness of our algorithm is proved by experimental results,where the average IOU is 0.82,and the average l2-norm distance is 9.It achieves real-time tracking with a speed of 34fps on PC with NVIDIA GTX 1080TI.
作者
杨勇
张轶
YANG Yong;ZHANG Yi(College of Computer Science,Sichuan University,Chengdu 610065;National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610065)
出处
《现代计算机》
2020年第14期75-82,共8页
Modern Computer
关键词
目标跟踪
目标检测
深度学习
模板更新
Object Tracking
Object Detection
Deep Learning
Template Update
