Trajectory anomalies serve as early indicators of potential issues and frequently provide valuable insights into event occurrence. Existing methods for detecting abnormal trajectories primarily focus on comparing the spatial characteristics of the trajectories. However, they fail to capture the temporal dimension’s pattern and evolution within the trajectory data, thereby inadequately identifying the behavioral inertia of the target group. A few detection methods that incorporate spatiotemporal features have also failed to adequately analyze the spatiotemporal sequence evolution information; consequently, detection methods that ignore temporal and spatial correlations are too one-sided. Recurrent neural networks (RNNs), especially gate recurrent unit (GRU) that design reset and update gate control units, process nonlinear sequence processing capabilities, enabling effective extraction and analysis of both temporal and spatial characteristics. However, the basic GRU network model has limited expressive power and may not be able to adequately capture complex sequence patterns and semantic information. To address the above issues, an abnormal trajectory detection method based on the improved GRU model is proposed in cloud computing in this paper. To enhance the anomaly detection ability and training efficiency of relevant models, strictly control the input of irrelevant features and improve the model fitting effect, an improved model combining the random forest algorithm and fully connected layer network is designed. The method deconstructs spatiotemporal semantics through reset and update gated units, while effectively capturing feature evolution information and target behavioral inertia by leveraging the integration of features and nonlinear mapping capabilities of the fully connected layer network. The experimental results based on the GeoLife GPS trajectory dataset indicate that the proposed approach improves both generalization ability by 1% and reduces training cost by 31.68%. This success do provides a practical solution for the task of anomaly trajectory detection.

中文翻译：

---

基于云计算深度时空序列分析的GRU模型的低成本高性能异常轨迹检测

轨迹异常可以作为潜在问题的早期指标，并经常为事件发生提供有价值的见解。现有的异常轨迹检测方法主要集中于比较轨迹的空间特征。然而，他们未能捕捉轨迹数据中时间维度的模式和演变，从而无法充分识别目标群体的行为惯性。一些结合时空特征的检测方法也未能充分分析时空序列演化信息；因此，忽略时间和空间相关性的检测方法过于片面。循环神经网络（RNN），特别是设计重置和更新门控制单元的门循环单元（GRU），具有非线性序列处理能力，能够有效提取和分析时间和空间特征。然而，基本的GRU网络模型的表达能力有限，可能无法充分捕获复杂的序列模式和语义信息。针对上述问题，本文提出一种基于改进的GRU模型的云计算异常轨迹检测方法。为了增强相关模型的异常检测能力和训练效率，严格控制不相关特征的输入，提高模型拟合效果，设计了随机森林算法与全连接层网络相结合的改进模型。该方法通过重置和更新门控单元解构时空语义，同时利用全连接层网络的特征集成和非线性映射能力有效捕获特征演化信息和目标行为惯性。基于GeoLife GPS轨迹数据集的实验结果表明，该方法泛化能力提高了1%，训练成本降低了31.68%。这一成功确实为异常轨迹检测任务提供了实用的解决方案。