引入全局上下文模块和高效注意力机制的车辆跟踪算法
DOI:
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

TP391

基金项目:

北京市长城学者培养计划


Vehicle tracking algorithm with global context module and efficient attention mechanism
Author:
Affiliation:

Fund Project:

Beijing Great Wall scholar training program

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    孪生全卷积神经网络目标跟踪算法(SiamFC)近些年成为车辆跟踪领域的研究热点。但该算法缺乏对目标车辆的深层特征提取和整体感知,在背景复杂、低分辨率、光照变化的情况下容易跟丢。该文提出使用深度残差网络ResNet50作为主干网络,根据跟踪模型特性,从剪裁特征图、调整网络总步长和嵌入高效通道注意力模块三方面对其进行优化,高效提取特征的同时增强模型的差异化认知,并在分支网络引入全局上下文模块(Non-local Network, NLNet),增强跟踪模型对目标车辆的整体感知。经实验证明,提出的算法在低分辨率、光照变化和复杂背景的情况下跟踪速度和鲁棒性显著提升。在VOT2018和OTB2015数据集中测试均能得到较好的跟踪结果, 与经典跟踪模型SiamFC相比,在OTB2015数据集中测试的跟踪精度提高了5.5%,跟踪成功率提高了2.7%, 跟踪速度提高了14%可达98FPS。

    Abstract:

    Fully-convolutional siamese neural network object tracking algorithm(SiamFC) has become a research hotspot in the field of vehicle tracking in recent years. However, this kind of algorithm lacks the ability of deep feature extraction and overall perception of the target vehicle, and is easy to be lost in the case of complex background、low resolution and illumination change. According to the characteristics of the vehicle tracking model, this paper propose to use the deep residual network as the backbone network, which is optimized from three aspects: clipping the feature map, adjusting the total step size of the network and embedding an efficient channel attention module, so as to extract features efficiently and enhance the differential recognition of the model. The Non-local network(Non-local Network, NLNet) is introduced into the branch network to enhance the overall perception of the target vehicle. The experimental results show that the robustness of the proposed algorithm is significantly improved in the case of low resolution, illumination change and complex background, and the real-time vehicle tracking is achieved. Good tracking results can be obtained from VOT2018 and OTB2015data sets, comparing with the classical tracking algorithm SiamFC, the tracking accuracy and tracking success rate of the algorithm proposed are improved 5.5% and 2.7%,the tracking speed is increased by 14% to 98FPS in the OTB2015data set.

    参考文献
    相似文献
    引证文献
引用本文

李畅,王一丁,孙芮,等. 引入全局上下文模块和高效注意力机制的车辆跟踪算法[J]. 科学技术与工程, 2022, 22(11): 4424-4433.
Li Chang, Wang Yiding, Sun Rui, et al. Vehicle tracking algorithm with global context module and efficient attention mechanism[J]. Science Technology and Engineering,2022,22(11):4424-4433.

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:2021-09-06
  • 最后修改日期:2022-01-21
  • 录用日期:2021-11-22
  • 在线发布日期: 2022-04-20
  • 出版日期:
×
律回春渐,新元肇启|《科学技术与工程》编辑部恭祝新岁!
亟待确认版面费归属稿件,敬请作者关注