【目的】针对现有轨迹异常检测方法未充分考虑道路网络约束的问题，提出一种有效检测出租车司机搭载乘客时可能存在的欺诈行为的轨迹异常检测算法。【方法】首先进行地图匹配，将轨迹数据映射到实际道路网络上，从而获取轨迹的路径序列。接着，使用两阶段聚类方法，先对匹配后的轨迹路径进行初始聚类，提取出核心路段并扩展，形成多个核心路径，再计算不同核心路径之间的相似性，并将相似性高的路径分配到同一簇，从而形成多个路径簇。最后，根据路径簇计算出通行代价阈值，再结合通行时间代价和通行路程代价计算每条轨迹的通行代价，将其与通行代价阈值进行比较，以此判断轨迹的异常性。【结果】在真实轨迹数据集上与传统异常检测方法进行对比，所提方法在异常轨迹检测上具有良好的效果，其运行时间明显低于STADCS方法，并且检测准确率比STADCS方法最高提升了9.03%； F₁分数值相较于Two Phase和ATDC方法也有明显提升，最高提升率分别为6.67%和9.45%。【结论】本文提出融合路网约束的检测方法，结合两阶段聚类和通行代价评估，有效提升检测准确性与效率，降低误判率。算法适用于复杂城市路网场景，为车辆轨迹数据挖掘及交通管理决策提供有力支持，在欺诈识别等领域具有重要应用价值。

[Objectives] To address the limitation that existing trajectory anomaly detection methods often fail to fully consider road network constraints, this study proposes a trajectory anomaly detection algorithm designed to effectively identify potential fraudulent behavior by taxi drivers during passenger pickup. [Methods] The algorithm first performs map matching, aligning trajectory data with the actual road network to obtain a sequence of path segments. Then, a two-stage clustering approach is applied: the matched trajectory paths are initially clustered to extract and expand core road segments, forming multiple core paths. Next, the algorithm calculates the similarity between different core paths and assigns highly similar paths to the same cluster, thereby generating multiple path clusters. Finally, a CostThreshold is computed based on each path cluster. The travel cost of each trajectory, calculated by combining travel time and distance costs, is compared against the corresponding CostThreshold to determine whether the trajectory is anomalous. [Results] Compared with traditional anomaly detection methods on real-world trajectory datasets, the proposed approach demonstrates superior performance in detecting anomalous trajectories. It achieves significantly lower runtime and improves detection accuracy by up to 9.03% compared to the STADCS method. The F₁ score also improves considerably compared to the Two-Phase and ATDC methods, with maximum gains of 6.67% and 9.45%, respectively. [Conclusions] This paper presents a detection method that integrates road network constraints with two-stage clustering and travel cost evaluation. The method enhances detection accuracy and efficiency while reducing the false positive rate. It is well-suited for complex urban road networks, offering valuable support for vehicle trajectory data mining and traffic management decision-making, with significant practical value in fraud detection and related fields.

利益冲突：Conflicts of Interest 所有作者声明不存在利益冲突。

All authors disclose no relevant conflicts of interest.