深浅特征融合的实时单目标行人跟踪
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

基金项目:

湖南省自然科学基金(2020JJ4201)


Real-time Single-object Pedestrian Tracking Based on Deep and Shallow Feature Fusion
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 增强出版
  • |
  • 文章评论
    摘要:

    单目标行人跟踪是计算机视觉目标跟踪领域最基础、也是研究最广泛的任务之一, 而目前大多数使用的相关滤波类算法和深度学习类算法则分别在跟踪精度和跟踪实时性上存在不足. 针对上述问题, 本文提出一种将目标图像的深浅特征融合的实时单目标行人跟踪方法. 算法利用卡尔曼滤波器预测目标位置, 通过计算四分颜色直方图提取目标的浅层颜色特征, 并获得预测相似性以判定预测的可靠性. 使用YOLOv4模型作为检测器, 提取目标深度特征并分别计算运动信息和外观信息的距离度量, 同时提取浅层颜色特征计算得到相似距离度量, 通过特征距离度量的加权融合对检测目标进行匹配与更新. 最后, 利用提出的轨迹更新策略协调预测和检测的调用关系, 达到准确性与实时性的平衡. 算法在OTB100和LaSOT数据集上进行了测试实验, 结果表明: 所提算法的跟踪准确率分别达到0.581和0.453, 在GPU上分别能达到33.64 FPS和35.32 FPS的跟踪速度, 满足实时跟踪的要求.

    Abstract:

    Single-object pedestrian tracking is one of the most basic and widely studied tasks in computer vision object tracking. However, most of the correlation filtering algorithms and deep learning algorithms currently used have insufficient tracking accuracy and real-time tracking performance. To solve the above problems, we propose a real-time single-object pedestrian tracking algorithm based on deep and shallow feature fusion. Firstly, this algorithm predicts the object location by Kalman filters and extracts the shallow color features of the object by calculating the four-part color histogram, and the prediction similarity is obtained to judge the reliability of prediction results. Then, YOLOv4 is used as a detector to extract deep features of the object and then calculate the distance metric of motion information and appearance information. Meanwhile, the shallow color features of the detection object are extracted to calculate the similarity distance metric, and the weighted fusion of the feature distance metric is employed to match the detection object and update the tracking trajectory. Finally, a trajectory updating strategy is put forward to coordinate the calling relationship between the prediction block and the detection block and to achieve a balance between tracking accuracy and speed. Testing experiments are conducted on the OTB100 and LaSOT datasets. The experimental results demonstrate that the tracking accuracy of the proposed algorithm on the above datasets reaches 0.581 and 0.453, respectively, and the tracking speed tested on GPU can achieve 33.64 FPS and 35.32 FPS, respectively, which meets the requirements of real-time tracking.

    参考文献
    相似文献
    引证文献
引用本文

王涛,王文格.深浅特征融合的实时单目标行人跟踪.计算机系统应用,,():1-8

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-11-12
  • 最后修改日期:2021-12-13
  • 录用日期:
  • 在线发布日期: 2022-05-30
  • 出版日期:
您是第位访问者
版权所有:中国科学院软件研究所 京ICP备05046678号-3
地址:北京海淀区中关村南四街4号 中科院软件园区 7号楼305房间,邮政编码:100190
电话:010-62661041 传真: Email:csa (a) iscas.ac.cn
技术支持:北京勤云科技发展有限公司

京公网安备 11040202500063号