ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2020, Vol. 57 ›› Issue (5): 996-1002.doi: 10.7544/issn1000-1239.2020.20190280

• 人工智能 • 上一篇    下一篇



  1. 1( 同济大学计算机科学与技术系 上海 201804);2( 嵌入式系统与服务计算教育部重点实验室(同济大学) 上海 201804) (
  • 出版日期: 2020-05-01
  • 基金资助: 

A Pedestrian Tracking Algorithm Based on Multi-Granularity Feature

Wang Ziye1, Miao Duoqian1,2, Zhao Cairong1,2 , Luo Sheng1, Wei Zhihua1,2   

  1. 1( Department of Computer Science and Technology, Tongji University, Shanghai 201804);2( Key Laboratory of Embedded System and Service Computing (Tongji University), Ministry of Education, Shanghai 201804)
  • Online: 2020-05-01
  • Supported by: 
    This work was supported by the National Key Research and Development Program of China (213) and the National Natural Science Foundation of China (61976158, 61673301).

摘要: 对于一些较为流行的应用,例如视频场景监控,对行人的长期有效跟踪是应用的基础.尽管对目标检测与跟踪的相关技术研究已经有了很长的历史,但是如何实时并较为准确地实现目标行人跟踪目前仍然是一个活跃的研究领域.基于多粒度的思想,提出了一种改进的行人跟踪算法,将卷积特征与底层颜色特征结合,对基于深度学习的跟踪算法GOTURN(generic object tracking using regression networks)得到的跟踪结果进行判断决策,结合目标检测对跟踪结果进行修正.实验结果表明:与单一的跟踪算法相比,多粒度决策的跟踪算法能够更加准确地对目标行人进行跟踪,可以显著提高跟踪精度.

关键词: 行人跟踪, 目标检测, 深度学习, 多粒度, 鲁棒性

Abstract: Recently in some popular applications, such as video scene surveillance, long-term effective pedestrian tracking is the basis of these applications. Although the related technology of target detection and target tracking have a long history, how to achieve real-time and accurate pedestrian tracking is still an active research field and needs to be solved. At present, most pedestrian tracking methods only use hand-designed features to track or only use deep learning to extract features, which are not good ways to represent the features of the target because the use of one single feature will restrict the expression of the features. Therefore, multi-granularity hierarchical features are used in this paper to achieve more stable pedestrian tracking. This paper proposes an improved pedestrian tracking algorithm. The algorithm adopts the idea of multi-granularity, combines convolutional feature with bottom color feature, makes decision on the tracking result obtained by GOTURN, a tracking algorithm based on deep learning, and modifies the tracking result with target detection. This paper uses Pascal VOC data set for model training, and uses OTB-100 and VOT 2015 data sets for testing. The experimental results show that the tracking algorithm based on multi-granularity decision can track target pedestrians more accurately than a single tracking algorithm and the tracking accuracy is improved obviously.

Key words: pedestrian tracking, object detection, deep learning, multiple-granularity, robustness