多特征融合的船舶轨迹相异度模型分析与优选

基于船舶自动识别系统轨迹，构建了船舶轨迹静态相异度模型、动态相异度模型以及组合相异度模型，包括轨迹起点和终点相异度模型、轨迹长度相异度模型、轨迹空间分布相异度模型、轨迹航速均值相异度模型、轨迹航向均值相异度模型、轨迹航速标准差相异度模型和轨迹航向标准差相异度模型；采用KNN分类算法进行轨迹分类，分析了单个相异度模型的有效性和时效性，对比了单个相异度模型和组合相异度模型下轨迹分类效果，研究了组合相异度模型中相异度模型的类别和权重对轨迹分类的影响；分别以内河航道和港口水域船舶轨迹进行试验。试验结果显示:在采用单个相异度的情况下，就分类效果而言，轨迹起点和终点相异度模型和轨迹航向均值相异度模型在内河航道和港口水域船舶轨迹分类效果均优于其他模型，而基于轨迹航速均值相异度模型和轨迹航速标准差相异度模型的轨迹分类效果最低，就分类效率而言，基于航速、航向均值和标准差的相异度模型耗时明显低于其他3个相异度模型；采用组合相异度进行轨迹分类，内河航道和港口水域船舶轨迹分类结果的基于精确率和召回率的宏平均值和微平均值均能接近99%；将组合相异度中相异度类别数由4个增加到7个，轨迹分类评估结果进一步得到提高。因此，单个相异度模型中以轨迹起点和终点相异度模型、轨迹航向均值相异度模型以及轨迹空间分布相异度模型分类效果最优且稳定，而轨迹空间分布相异度模型和轨迹长度相异度模型耗时明显高于其他方式，各相异度模型在不同场景中的适应性基本相似，通过增加组合相异度中相异度类别能够提高轨迹识别效果。 

Analysis and optimization of ship trajectory dissimilarity models based on multi-feature fusion

Based on ship automatic identification system (AIS) trajectories, static dissimilarity models, dynamic dissimilarity models, and a combined dissimilarity model of ship trajectories were constructed, including the following dissimilarity models: trajectory departure and destination, trajectory length, trajectory spatial distribution, trajectory mean speed, trajectory mean course, trajectory speed standard deviation, and trajectory course standard deviation. Trajectories were classified using the KNN classification algorithm, the effectivenesses and efficiencies of each single dissimilarity model were analyzed, the effect of trajectory classification under different unique dissimilarity models and the combined dissimilarity model were compared, and the influence of the categories and weights of dissimilarity models on trajectory classification in the combined dissimilarity model was studied. Experiments were conducted using ship trajectories in inland waterways and port waters. Experimental results show that under the condition of adopting a single dissimilarity, in terms of the classification effect, the ship trajectory classification based on the dissimilarity model of trajectory departure and destination and the dissimilarity model of trajectory mean course is better than that using other dissimilarity models in inland waterways and port waters, whereas the trajectory classification effect based on the dissimilarity model of trajectory mean speed and the dissimilarity model of trajectory speed standard deviation is worse. In terms of classification efficiency, the time consumed by the dissimilarity models based on mean value and standard deviation is significantly lower than that of the other dissimilarity models. Through the analysis and optimization of trajectory dissimilarity models based on the trajectory classification results of the KNN classification algorithm, when the trajectory classification is conducted using the combined dissimilarity model, macro and micro averages based on accuracy and recall of ship trajectory classification results in the inland waterway and port waters can both reach 99%; moreover, by increasing the number of dissimilarity categories in the combined dissimilarity from 4 to 7, the evaluation result of trajectory classification is further improved. Therefore, in the single dissimilarity model, the classification effects of the dissimilarity model of trajectory departure and destination, the dissimilarity model of trajectory mean course, and the dissimilarity model of trajectory spatial distribution are optimal and stable, whereas the time consumption of the dissimilarity model of trajectory spatial distribution and the dissimilarity model of trajectory length are significantly higher than those of other models. The adaptabilities of each dissimilarity are similar in different scenarios. By increasing the dissimilarity category in the combined dissimilarity model, the trajectory recognition effect can be improved. 3 tabs, 12 figs, 30 refs.