计算机应用 ›› 2021, Vol. 41 ›› Issue (2): 523-529.DOI: 10.11772/j.issn.1001-9081.2020060762

所属专题: 多媒体计算与计算机仿真

• 多媒体计算与计算机仿真 • 上一篇    下一篇

基于孪生区域候选网络的无人机指定目标跟踪

钟莎, 黄玉清   

  1. 西南科技大学 信息工程学院, 四川 绵阳 621010
  • 收稿日期:2020-06-08 修回日期:2020-09-18 出版日期:2021-02-10 发布日期:2020-12-17
  • 通讯作者: 黄玉清
  • 作者简介:钟莎(1996-),女,四川成都人,硕士研究生,主要研究方向:图像处理、机器视觉;黄玉清(1962-),女,四川绵阳人,教授,硕士,主要研究方向:图像处理、机器视觉、智能技术。
  • 基金资助:
    国家自然科学基金资助项目(61673220)。

Specified object tracking of unmanned aerial vehicle based on Siamese region proposal network

ZHONG Sha, HUANG Yuqing   

  1. School of Information Engineering, Southwest University of Science and Technology, Mianyang Sichuan 621010, China
  • Received:2020-06-08 Revised:2020-09-18 Online:2021-02-10 Published:2020-12-17
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61673220).

摘要: 基于孪生网络的目标跟踪目前取得了阶段性进展,即克服了孪生网络的空间不变性在深度网络中的限制,然而其仍存在外观变化、尺度变化、遮挡等因素影响跟踪性能。针对无人机(UAV)指定目标跟踪中的目标尺度变化大、目标运动模糊及目标尺度小等问题,提出了基于孪生区域候选注意力机制网络的跟踪算法Attention-SiamRPN+。首先,采用改进的深度残差网络ResNet-50作为特征提取器来提取特征;接着,使用通道注意力机制模块筛选残差网络提取出的不同通道特征图的语义属性,并重新为不同通道特征分配相应权值;然后,两个区域候选网络(RPN)进行分层融合,而RPN模块包括特征图的逐通道深度互相关、正负样本分类和边界框回归;最后框选出目标位置。在VOT2018平台上进行测试,所提算法的准确率和预期平均重叠率(EAO)分别为59.4%和39.5%;在OTB2015平台上采用一次通过评估模式进行实验,该算法的成功率和精度分别为68.7%和89.4%。实验结果表明所提算法的评估结果优于近年优秀的三种相关滤波跟踪算法和孪生网络跟踪算法,且该算法应用于UAV指定目标的跟踪上时具有良好的鲁棒性和实时处理速度。

关键词: 孪生网络, 深度残差网络, 注意力机制, 无人机, 目标跟踪

Abstract: Object tracking based on Siamese network has made some progresses, that is it overcomes the limitation of the spatial invariance of Siamese network in the deep network. However, there are still factors such as appearance changes, scale changes, and occlusions that affect tracking performance. Focusing on the problems of large changes in object scale, object motion blur and small scale of object in the specified object tracking of Unmanned Aerial Vehicles (UAV), a new tracking algorithm was proposed based on the Siamese region proposal attention mechanism network, namely Attention-SiamRPN+. Firstly, an improved deep residual network ResNet-50 was employed as a feature extractor to extract feature maps. Secondly, the channel attention mechanism module was used to filter the semantic information of different channel feature maps extracted by the residual network, and the corresponding weights to different channel features were reassigned. Thirdly, a hierarchical fusion of two Region Proposal Networks (RPN) was applied. The RPN module was consisted of channel-by-channel deep cross-correlation of feature maps, classification of positive and negative samples and bounding box regression. Finally, the box of the object position was selected. In the test on the VOT2018 platform, the proposed algorithm had the accuracy of 59.4% and the Expected Average Overlap (EAO) of 39.5%. In the experiment with one-pass evaluation mode on the OTB2015 platform, the algorithm had the success rate and precision of 68.7% and 89.4% respectively. Experimental results show that the evaluation results of the proposed algorithm are better than the results of three excellent correlation filtering tracking and Siamese network tracking algorithms in recent years, and the proposed algorithm has good robustness and real-time processing speed when applying to the tracking of specified objects of UAV.

Key words: Siamese network, deep residual network, attention mechanism, Unmanned Aerial Vehicle (UAV), object tracking

中图分类号: