计算机工程 ›› 2019, Vol. 45 ›› Issue (9): 216-221.doi: 10.19678/j.issn.1000-3428.0053937

• 图形图像处理 • 上一篇    下一篇

一种用于单目标跟踪的锚框掩码孪生RPN模型

李明杰, 冯有前, 尹忠海, 周诚, 董方昊   

  1. 空军工程大学 基础部, 西安 710051
  • 收稿日期:2019-02-19 修回日期:2019-03-25 出版日期:2019-09-15 发布日期:2019-09-03
  • 作者简介:李明杰(1995-),男,硕士研究生,主研方向为计算机视觉、目标检测;冯有前,教授、博士生导师;尹忠海,教授;周诚,博士研究生;董方昊,硕士研究生。
  • 基金项目:
    国家自然科学基金(61472443)。

An Anchor Mask Siamese RPN Model for Single Target Tracking

LI Mingjie, FENG Youqian, YIN Zhonghai, ZHOU Cheng, DONG Fanghao   

  1. Department of Sciences, Air Force Engineering University, Xi'an 710051, China
  • Received:2019-02-19 Revised:2019-03-25 Online:2019-09-15 Published:2019-09-03

摘要: 针对孪生区域候选网络(RPN)易受干扰且目标丢失后无法跟踪的问题,引入锚框掩码网络机制,设计一种新型孪生RPN模型。设置多尺度模板图片,并将其与目标图片进行卷积操作,实现全图检测以避免目标丢失。通过对前三帧图片的IOU热度图进行学习,预测连续帧目标锚框掩码,简化计算并排除其他目标干扰。在VOT2016和OTB100数据集中的实验结果显示,该模型对VOT2016数据集检测帧率达到24.6 frame/s,预期平均覆盖率为0.344 5,对OTB100数据集的检测准确率和成功率分别为0.862和0.642。基于摄像头采集数据的目标丢失及干扰测试表明,该模型具有良好的抗干扰性与实时性。

关键词: 孪生区域候选网络, 锚框掩码, 锚框掩码网络, 多尺度变换, 目标跟踪

Abstract: To address the problem that the Siamese Region Proposal Network(RPN) is susceptible to interference and cannot be tracked after the target is lost,this paper introduces the anchor mask network mechanism to design a new Siamese RPN model.The model sets the multi-scale template images and convolves them with the target image to achieve full-image detection and avoid target loss.By learning the IOU hot maps of the first three frames,the target anchor mask is predicted in the continuous frames to simplify the calculation and exclude other target interference.The experimental results in the VOT2016 and OTB100 datasets show that the model has a detection rate of 24.6 frame/s and an expected average overlap of 0.344 5 for the VOT2016 dataset,and a precision of 0.862 and a success rate of 0.642 for the OTB100 dataset.The target loss and interference tests are carried out on the data collected by the camera.The results show that the model has good anti-interference and real-time performance.

Key words: Siamese Region Proposal Network(RPN), anchor mask, anchor mask network, multi-scale transformation, target tracking

中图分类号: