顾及运动行为特征变化的船舶轨迹分类模型  

A ship trajectory classification model considering changes in characteristics of motion behavior

作　　者：张鹏鑫 李连营[1] 杨敏[1] 安晓亚[2,3] 许小兰 ZHANG Pengrin;LI Lianying;YANG Min;AN Xiaoya;XU Xiaolan(School of Resource and Environment Science,Wuhan University,Wuhan 430079,China;Xi'an Research Institute of Surveying and Mapping,Xi'an 710054,China;State Key Laboratory of Geo-information Engineering,Xi'an 710054,China;School of Urban Design,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学资源与环境科学学院,武汉430079 [2]西安测绘研究所,西安710054 [3]地理信息工程国家重点实验室,西安710054 [4]武汉大学城市设计学院,武汉430079

出　　处：《测绘科学》2023年第5期25-34,共10页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(42271458,41871377);基础加强计划重点基础研究项目(2020-JCJQ-ZD-087)。

摘　　要：针对传统轨迹分类方法依赖人工定义特征和设置参数,以及深度学习支持下的轨迹分类模型大多数会将轨迹数据组织为栅格图像,导致轨迹点隐含的运动行为特征及其在时间序列上的变化特点难以有效顾及的问题,该文提出一种卷积神经网络(CNN)和长短期记忆(LSTM)网络组合的船舶轨迹分类模型。在计算轨迹点隐含运动属性并构建船舶运动属性序列的基础上,首先利用CNN提取局部区域运动属性的高层次特征,然后按时间次序输入LSTM网络进一步分析运动过程的行为变化,从而实现船舶轨迹的分类识别。采用真实的轨迹数据进行试验,结果表明:CNN和LSTM组合的分类模型有效顾及了船舶的行为特征变化,其分类准确率达到87.1%,优于单一CNN和LSTM模型的分类准确率。Aiming at the problem that traditional trajectory classification methods relied on the manually defined features and setting of parameters,and most of the trajectory classification models supported by deep learning converted trajectory data into raster images,which produced the difficulty of effectively taking into account the characteristics of motion behavior implied by trajectory points and their variation characteristics in time series,this paper proposed a ship trajectory classification model combining Convolutional Neural Network(CNN)and Long Short-Term Memory(LSTM)network.Based on calculating the implied motion attributes of the trajectory points and constructing a sequence of motion attributes,the CNN model was firstly used to extract the high-level features of the motion attributes in the local areas.Then,the outputs of the CNN-based model was organized as the inputs of the LSTM network to further analyze the behavioral changes of the moving process,so as to recognize the classes of ship trajectories.Experiments were conducted using real trajectory data,and the results showed that the classification model combining CNN and LSTM effectively took into account the changes in characteristics of motion behavior of ships,and its classification accuracy reached 87.1%,which was better than the classification accuracy of single CNN and LSTM models.

关 键 词：船舶轨迹 分类 卷积神经网络 长短期记忆网络 

分 类 号：P228[天文地球—大地测量学与测量工程]