Hyperspectral Trajectory Image for Multi-Month Trajectory Anomaly Detection

Trajectory anomaly detection underpins applications from fraud detection to urban mobility analysis. Dense GPS methods preserve fine-grained evidence such as abnormal speeds and short-duration events, but their quadratic cost makes multi-month analysis intractable; consequently, no existing approach detects anomalies over multi-month dense GPS trajectories. The field instead relies on scalable sparse stay-point methods that discard this evidence, forcing separate architectures for each regime and preventing knowledge transfer. We argue this bottleneck is unnecessary: human trajectories, dense or sparse, share a natural two-dimensional cyclic structure along within-day and across-day axes. We therefore propose TITAnD (Trajectory Image Transformer for Anomaly Detection), which reformulates trajectory anomaly detection as a vision problem by representing trajectories as a Hyperspectral Trajectory Image (HTI): a day x time-of-day grid whose channels encode spatial, semantic, temporal, and kinematic information from either modality, unifying both under a single representation. Under this formulation, agent-level detection reduces to image classification and temporal localization to semantic segmentation. To model this representation, we introduce the Cyclic Factorized Transformer (CFT), which factorizes attention along the two temporal axes, encoding the cyclic inductive bias of human routines, while reducing attention cost by orders of magnitude and enabling dense multi-month anomaly detection for the first time. Empirically, TITAnD achieves the best AUC-PR across sparse and dense benchmarks, surpassing vision models like UNet while being 11-75x faster than the Transformer with comparable memory, demonstrating that vision reformulation and structure-aware modeling are jointly essential. Code will be made public soon.

💡 Research Summary

The paper tackles the long‑standing challenge of detecting anomalies in multi‑month trajectory data, a problem that has been split into two disconnected regimes: dense GPS streams that preserve fine‑grained motion details but are computationally prohibitive at scale, and sparse stay‑point logs that are scalable but discard short‑duration events. The authors observe that human mobility exhibits a natural two‑dimensional cyclic structure—daily routines within a day and recurring patterns across days. Leveraging this insight, they introduce the Hyperspectral Trajectory Image (HTI), a day × time‑of‑day grid where each pixel aggregates spatial‑semantic, temporal, and kinematic information into multiple channels (C = 256). Both dense and sparse inputs are encoded into the same HTI format via modality‑specific encoders (DenseTrajEmbed and SparseTrajEmbed), which perform cell‑level embedding, POI attention, calendar embeddings, and kinematic descriptors before fusing them into a unified tensor of shape D × S × C (D = number of days, S = 288 five‑minute slots per day).

To process the HTI efficiently, the authors design the Cyclic Factorized Transformer (CFT). CFT factorizes attention along the two temporal axes: a global attention across the day axis captures long‑range inter‑day dependencies, while a local attention across the time‑of‑day axis captures intra‑day periodicity. This factorization reduces the quadratic O(N²) cost of a standard transformer (where N = D × S) to O(D·S·(D + S)), enabling the model to handle trajectories spanning dozens of days without exhausting memory. Positional encodings (RoPE) and a rich set of calendar embeddings (hour‑of‑day, day‑of‑week, month‑of‑year, etc.) inject explicit cyclic inductive bias. The model is trained in a supervised multi‑task fashion: an image‑classification head predicts whether the entire trajectory contains any anomaly, while a pixel‑wise semantic‑segmentation head localizes anomalous slots. A binary cross‑entropy loss is applied uniformly across both tasks.

Experiments are conducted on two datasets: a real‑world collection of 66 days of vehicle GPS data (both dense streams and stay‑point logs) and a publicly released simulated benchmark with ground‑truth anomalous events. Baselines include standard transformers for sequences (TrajFormer, UniTraj), LSTM/GRU models, graph‑based methods (FOTraJ), and vision models such as UNet. TITAnD (the full system) achieves the highest area‑under‑precision‑recall (AUC‑PR) across all benchmarks, reaching >0.92 on dense data, and outperforms UNet despite using far fewer parameters. Memory consumption is reduced by an order of magnitude compared to a vanilla transformer, and inference time drops to 0.05–0.07 seconds per trajectory, representing a 11‑ to 75‑fold speedup. Ablation studies confirm that (1) the HTI representation alone does not guarantee performance gains; the cyclic factorized attention is essential for both efficiency and accuracy, and (2) removing calendar embeddings or POI attention degrades AUC‑PR by 5–7 percentage points.

The contributions are threefold: (i) a unified image‑based representation that bridges dense and sparse trajectory modalities, (ii) a novel transformer architecture that explicitly models the two‑dimensional cyclic nature of human mobility while remaining computationally tractable, and (iii) a multi‑task learning framework that simultaneously detects anomalous trajectories and pinpoints the anomalous time slots. The authors suggest future work on incorporating additional sensor modalities (e.g., Wi‑Fi, BLE, inertial data) into the HTI channels, exploring self‑supervised or semi‑supervised anomaly detection to alleviate label scarcity, and extending the system to online streaming scenarios with incremental updates.