计算机工程

• 图形图像处理 • 上一篇    下一篇

基于置信图特性的改进时空上下文目标跟踪

张雷,于凤芹   

  1. (江南大学 物联网工程学院,江苏 无锡 214122)
  • 收稿日期:2015-07-30 出版日期:2016-08-15 发布日期:2016-08-15
  • 作者简介:张雷(1990-),男,硕士研究生,主研方向为视频图像处理;于凤芹,教授、博士。

Improved Object Tracking via Spatial-temporal Context Based on Confidence Map Property

ZHANG Lei,YU Fengqin   

  1. (School of Internet of Things Engineering,Jiangnan University,Wuxi,Jiangsu 214122,China)
  • Received:2015-07-30 Online:2016-08-15 Published:2016-08-15

摘要: 针对时空上下文目标跟踪算法在目标处于遮挡情况下容易产生漂移的问题,基于置信图特性,提出一种改进的时空上下文目标跟踪算法。利用3种子块的特征组合表达目标外观,将置信图中多个峰值点的对应区域作为候选区域,从而提取候选区域的目标特征并找到与目标模板最相似的区域。通过连续蒙特卡洛采样得到最优目标区域,并根据子块遮挡比例自适应调节时空上下文学习率以降低遮挡的影响。仿真实验结果表明,与时空上下文目标跟踪算法和压缩跟踪算法相比,在目标快速移动或发生遮挡时,改进算法仍能较准确地跟踪目标。

关键词: 目标跟踪, 时空上下文, 峰值点, 子块组合, 遮挡比例, 学习率

Abstract: Object tracking algorithm via Spatial-temporal Context(STC) suffers from drift when the object is under occlusion,so this paper proposes an improved object tracking algorithm via STC based on confidence map property.Three types of combinations of sub-block features are used to represent the object.The regions corresponding to the multiple peak points in confidence map are regarded as the candidate regions.Then features of the regions are extracted to match the object template to find out which is most similar with the object.Sequential Monte Carlo method is used to obtain the best object region.Finally,the learning rate of STC is adjusted through object sub-block occlusion ratio to reduce the effect of occlusion.Simulation results show that the proposed algorithm can still track the object accurately in case of rapid movement or occlusion,compared with the tracking algorithm via STC and the compressive tracking algorithm.

Key words: object tracking, Spatial-temporal Context(STC), peak point, combination of sub-block, occlusion ratio, learning rate

中图分类号: