生物多样性 ›› 2024, Vol. 32 ›› Issue (3): 23409.  DOI: 10.17520/biods.2023409  cstr: 32101.14.biods.2023409

• 技术与方法 • 上一篇    下一篇

基于注意力机制融合多特征的东北虎个体自动跟踪方法

许群(), 谢永华*()()   

  1. 东北林业大学计算机与控制工程学院, 哈尔滨 150040
  • 收稿日期:2023-10-31 接受日期:2024-01-29 出版日期:2024-03-20 发布日期:2024-03-24
  • 通讯作者: *E-mail: zdhxyh@163.com
  • 基金资助:
    黑龙江省自然科学基金(LH202C034)

Automatic individual tracking method of Amur tiger based on attention mechanism fusion of multiple features

Qun Xu(), Yonghua Xie*()()   

  1. College of Computer and Control Engineering, Northeast Forestry University, Harbin 150040
  • Received:2023-10-31 Accepted:2024-01-29 Online:2024-03-20 Published:2024-03-24
  • Contact: *E-mail: zdhxyh@163.com

摘要:

跟踪野生动物可以了解它们的生态习性和种群动态。自动化的、高效准确的目标跟踪算法对于野生动物保护具有重要意义。由于东北虎(Panthera tigris altaica)生存环境复杂, 行动方式隐蔽且具有快速运动的特点, 拍摄到的图像可能存在运动模糊问题, 难以捕获清晰稳定的画面。因此, 监测东北虎种群的难点在于实现自动准确地跟踪东北虎个体。在实际环境中, 由于光照变化、遮挡、相似性干扰等问题的存在, 导致错误跟踪东北虎个体。针对这些问题, 本文提出了一种基于注意力特征融合的孪生网络跟踪框架, 旨在实现对实际复杂场景中东北虎个体的准确跟踪。基于孪生网络的视觉目标跟踪框架将目标跟踪视为相似性学习问题, 本文对传统基于孪生网络的跟踪架构进行改进, 将注意力特征融合嵌入主干网络ResNet50中进行特征提取。为了增强模型对跟踪场景中发生极端尺度变化的东北虎个体的关注度, 本文在注意力特征融合模块中引入了多尺度通道注意力机制, 以适应不同的东北虎个体状态和环境变化。实验结果表明, 本文的方法与当前的先进跟踪器相比取得了更优的跟踪性能, 跟踪成功率和精确度分别达到了72.5%和93.9%, 相比基准跟踪算法分别提高了4.1%和2.3%, 证明本文的跟踪方法可以在复杂的实际场景下为自动高效地监测东北虎提供一种有效的方案。

关键词: 东北虎, 单目标跟踪, 孪生网络, 特征融合, 注意力机制

Abstract

Background: Tracking wild animals sheds light on their ecology, behavior, and population dynamics. Developing an automated, effective, and precise target tracking system is crucial for the conservation of wild animals. Monitoring the Amur tiger (Panthera tigris altaica) population requires properly and automatically tracking individual tigers, which is challenging because of their camouflaged and fast movements in complex habitats. Individual Amur tigers are tracked incorrectly in the real world because of factors like similarity interference, occlusion, and illumination variation.
Methods: In order to accurately track Amur tigers in complex real-world circumstances, we have suggested a Siamese network tracking framework based on attention feature fusion. Through the incorporation of the attention feature fusion into the backbone network ResNet50, we have enhanced the conventional Siamese-based tracking architecture. With the addition of a multi-scale channel attention module, the system was better able to comprehend global contextual information and adjust to the varying environmental conditions and individual states of Amur tigers.
Results: The suggested approach was compared with five advanced algorithms, SiamFC, SiamRPN++, SiamCAR, SiamBAN and SiamGAT, on the Amur tiger target tracking dataset. The algorithm proposed in this paper achieved a tracking success rate of 72.5% and a precision of 93.9%, outperforming the five existing algorithms. It showed improvements of 4.1% and 2.3% compared to the baseline tracking algorithm. At the same time, the suggested approach performed better in tracking when faced with six distinct tracking problems in the Amur tiger’s complex environment.
Conclusion: This strategy significantly enhances the success rate and precision of tracking individual Amur tigers. The method in this paper is more suitable for the actual scene under the premise of using computer vision technology to monitor wild animals, and it proves that the tracking method in this paper can provide an effective scheme for automatic and efficient monitoring of Amur tigers in a complex actual scene.

Key words: Amur tiger, single object tracking, Siamese network, feature fusion, attention mechanism