计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (7): 121-129.DOI: 10.3778/j.issn.1002-8331.1912-0405

• 模式识别与人工智能 • 上一篇    下一篇

深度学习及时空约束的行人跟踪算法研究

唐国智,李顶根   

  1. 华中科技大学 能源与动力工程学院,武汉 430074
  • 出版日期:2021-04-01 发布日期:2021-04-02

Research on Pedestrian Tracking Algorithms with Deep Learning and Space-Time Constraints

TANG Guozhi, LI Dinggen   

  1. School of Energy and Power Engineering, Huazhong University of Science and Technology, Wuhan 430074, China
  • Online:2021-04-01 Published:2021-04-02

摘要:

由于以往的行人跟踪方法大部分不能有效地解决目标被遮挡后以及目标尺寸变化再跟踪的问题,所以引入了深度学习的方法,但是经实验发现单纯使用深度学习行人跟踪会因行人检测部分的误差而出现整体的跟踪准确率不高的问题。提出了一种基于深度学习和时空约束后处理的行人跟踪方法,深度学习的行人检测部分采用了根据实际应用场景优化过的SSD算法,行人匹配部分采用了一种计算交叉输入领域差异然后进行块总结的方法,最后进行时空约束的后处理。在OTB数据集上做实验,与传统跟踪算法以及单纯深度学习算法进行了对比。

关键词: 深度学习, 行人检测, 行人匹配, 时空约束

Abstract:

Because most of the previous pedestrian tracking methods cannot effectively solve the problem of target occlusion and target size change and re-tracking, the deep learning method is introduced, but it is found through experiments that the use of deep learning pedestrian tracking will be due to the error of pedestrian detection. And the overall tracking accuracy is not high. A pedestrian tracking method based on deep learning and spatiotemporal constraint post-processing is proposed. The pedestrian detection part of deep learning adopts SSD algorithm. The pedestrian matching part adopts a method of calculating cross-input field difference and then performing block summarization, finally, the post-processing of space-time constraints is carried out. Experiments on the OTB dataset are compared with the traditional tracking algorithm and the pure deep learning algorithm.

Key words: deep learning, pedestrian detection, pedestrian matching, space-time constrain