基于多源数据融合与GMM时空基线的车辆异常轨迹检测与风险预警平台设计与实现

Design and Implementation of a Vehicle Abnormal-Trajectory Detection and Risk-Warning Platform Based on Multi-Source Data Fusion and a GMM Spatiotemporal Baseline

摘要: 针对车辆轨迹数据来源异构、时空连续性不足以及固定规则难以刻画不同车辆正常出行规律的问题，本文设计并实现一套基于多源数据融合与个性化时空基线的车辆异常轨迹检测与风险预警平台。平台以车辆GPS定位、道路卡口识别以及停车场出入口记录等多源轨迹数据为基础，通过坐标标准化、时间对齐、插值补全和卡尔曼滤波完成异构轨迹融合；进一步提取停留区域、常驻区域和高频路线等车辆出行特征，并将空间位置与小时、星期周期特征联合建模，采用高斯混合模型(GMM)建立车辆历史正常出行时空基线。在此基础上，以轨迹点在基线模型中的负对数似然作为异常分数，依据历史正常样本异常分数分布设置阈值，并将异常程度映射为低、中、高风险及P3、P2、P1三级预警。平台采用Streamlit构建Web交互界面，结合Folium及地图瓦片服务实现轨迹地图、异常点、热力图、风险等级分布和预警记录的可视化展示。考虑真实车辆轨迹具有位置敏感性和业务数据安全要求，本文系统验证全部采用程序生成的虚拟多源轨迹数据，模拟结果仅用于验证“数据生成–轨迹融合–基线建模–异常检测–风险预警–可视化展示”的端到端流程，不用于表征真实业务场景下的检测精度。系统运行表明，所设计平台能够完成从多源轨迹处理到分级预警输出的完整功能闭环，可为车辆异常轨迹分析、风险研判及相关教学科研和原型系统构建提供工程化参考。

Abstract: To address the heterogeneity of vehicle trajectory data sources, insufficient spatiotemporal continuity, and the limited ability of fixed-rule methods to characterize the normal travel patterns of different vehicles, this paper designs and implements a vehicle anomalous trajectory detection and risk early-warning platform based on multi-source data fusion and personalized spatiotemporal baselines. The platform integrates multi-source trajectory data, including vehicle GPS positioning data, road checkpoint recognition records, and parking-lot entrance and exit records. Coordinate standardization, temporal alignment, interpolation, and Kalman filtering are applied to fuse heterogeneous trajectory data and construct continuous trajectories. Vehicle mobility features, such as stay regions, frequently visited regions, and high-frequency routes, are then extracted. Spatial locations are jointly modeled with cyclic hour-of-day and day-of-week features, and a Gaussian Mixture Model (GMM) is employed to establish a personalized spatiotemporal baseline of historical normal travel behavior. The negative log-likelihood of each trajectory point under the baseline model is used as the anomaly score, and the anomaly threshold is determined according to the distribution of anomaly scores derived from historical normal samples. The degree of anomaly is further mapped to low-, medium-, and high-risk levels and to P3, P2, and P1 warning levels. A Streamlit-based Web interface, together with Folium and map tile services, is used to visualize vehicle trajectories, anomalous points, heat maps, risk-level distributions, and warning records. Considering the sensitivity of real vehicle location data and the security requirements of operational data, all system validation in this study is conducted using programmatically generated synthetic multi-source trajectory data. The simulation results are intended only to verify the end-to-end workflow of data generation, trajectory fusion, baseline modeling, anomaly detection, risk early warning, and visualization, rather than to represent detection accuracy in real-world operational scenarios. The system results demonstrate that the proposed platform can implement a complete functional workflow from multi-source trajectory processing to hierarchical warning output, providing an engineering reference for vehicle anomalous trajectory analysis, risk assessment, related teaching and research, and prototype system development.