Abstract:

Traditional visual ship tracking task mainly focuses on single target tracking, and thus less attention is paid to multi-ship tracking task. To address the issue, a novel ship tracking framework based on multi-dimension feature fusion mechanism and scale change estimation is proposed. The ship target position is firstly estimated with multi-dimension features by finding the maximum response between the training sample and the input maritime image, and its scale is then estimated with a scale filter. The experimental results show that average ship tracking error in different maritime traffic scenarios are less than 10 pixels, which verifies the effectiveness and reliability of the proposed framework.

Key words: multi-dimension feature fusion, scale variation, correlation filtering mechanism, multi-target ship tracking, smart ship

摘要：

传统的船舶视觉跟踪任务主要集中于单目标船舶跟踪，对多目标船舶跟踪研究相对较少。为解决该问题，提出一种多维特征融合机制和尺度变化估计的多目标船舶跟踪框架，该框架引入位置滤波器对输入的船舶训练样本进行学习，并将其应用于待跟踪的船舶图片序列，通过寻找最大响应的方法判定图像中的船舶位置。在此基础上，构建船舶尺度估计滤波器以确定待跟踪船舶的图像尺寸。通过和中值流跟踪算法和多示例学习跟踪算法对比分析，实验结果表明不同海事交通场景下的船舶跟踪误差均小于10像素，验证了算法的有效性和可靠性。

关键词: 多维特征融合机制, 尺度变化, 相关滤波机制, 多目标船舶跟踪, 智能船舶