周羽, 黄亮, 周春辉, 等. 基于轨迹特征图像深度学习的船舶时空行为分类识别方法[J]. 中国舰船研究, 2024, 19(X): 1–11. DOI: 10.19693/j.issn.1673-3185.03939
引用本文: 周羽, 黄亮, 周春辉, 等. 基于轨迹特征图像深度学习的船舶时空行为分类识别方法[J]. 中国舰船研究, 2024, 19(X): 1–11. DOI: 10.19693/j.issn.1673-3185.03939
shu
ZHOU Y, HUANG L, ZHOU C H, et al. Classification and recognition of spatio-temporal behavior of ships based on deep learning of trajectory feature images[J]. Chinese Journal of Ship Research, 2024, 19(X): 1–11 (in Chinese). DOI: 10.19693/j.issn.1673-3185.03939
Citation: ZHOU Y, HUANG L, ZHOU C H, et al. Classification and recognition of spatio-temporal behavior of ships based on deep learning of trajectory feature images[J]. Chinese Journal of Ship Research, 2024, 19(X): 1–11 (in Chinese). DOI: 10.19693/j.issn.1673-3185.03939
shu

基于轨迹特征图像深度学习的船舶时空行为分类识别方法

Classification and recognition of spatio-temporal behavior of ships based on deep learning of trajectory feature images

    摘要
  • 摘要:
    目的 针对现有船舶行为识别方法在处理大规模海上船舶轨迹数据时存在效率低、准确性差等问题，提出一种基于轨迹特征图像建模与深度学习的船舶行为识别及分类方法，旨在快速、高效识别和分类船舶行为模式。
    方法 考虑船舶轨迹多尺度特征，提出船舶轨迹的自适应网格化处理方法，构建航速、加速度、航向、转向率和轨迹点密度等显著特征的视觉编码模型，实现船舶轨迹特征图像的样本生成和增强处理，进而采用卷积神经网络构建船舶行为分类识别模型，对不同船舶行为的轨迹图像样本集进行训练和验证。
    结果 实验结果表明，航速、转向率和轨迹点密度是区分直航、转向、机动徘徊、靠泊和锚泊等8种行为的最佳特征组合，基于轨迹特征图像的深度学习模型能显著提高船舶行为识别的质量和精度：召回率为90.99%，精确度为91.23%，F1分数为91.11%，准确率达到91.22%。
    结论 该方法可有效识别不同尺度轨迹数据的船舶行为，开展区域船舶行为的自动分类识别，结果可为水上交通智能管控提供决策支撑。

     

    Abstract:
    Objective To address the issues of low efficiency and inaccurate ship behavior recognition when handling large-scale ship trajectory data, a method for recognizing and classifying ship behaviour based on trajectory feature image modelling and deep learning is proposed.
    Method This research constructs a visual coding model for the salient features, including speed, acceleration, heading, steering rate and trajectory point density. It also realizes sample generation and enhancement processing of ship trajectory feature images. All of these are done while taking into account the multi-scale features of ship trajectory.
    Results Based on the trajectory feature images, the deep learning model significantly improves the quality and accuracy of ship behavior recognition, with a recall rate of 90.99%, precision rate of 91.23%, and F1 score of 91.11%, which translates to an accuracy rate of 91.22%.The experimental results indicate that the speed, steering rate, and trajectory point density are the best feature combinations to distinguish the eight behaviors, such as straight ahead, steering, maneuvering loitering, berthing, and anchoring.
    Conclusion The approach can successfully detect ship behaviors at various trajectory data scales, perform automatic ship behavior categorization and identification, and produce outcomes that may help make decisions for intelligent water traffic control.

     

    • HTML全文
    • 参考文献(33)
    • 相关文章
    • 施引文献
  • 资源附件(0)

/

返回文章
返回