Parallel recurrent neural network with transformer for anomalous trajectory detection

Dawen Xia, Yunsong Li, Yuce Ao, Xiaoduo Wei, Yan Chen, Yang Hu, Yantao Li, Huaqing Li

Abstract: Anomalous trajectory detection within urban road traffic networks is crucial for identifying operational vehicle fraud in intelligent transportation systems. However, most existing approaches are limited to detecting anomalous trajectories solely based on the same original point, neglecting the extraction of spatiotemporal features and contextual information embedded in trajectory data. To address these limitations, a Parallel Recurrent Neural Network with Transformer (PRNNT) model is proposed for anomalous trajectory detection. Specifically, the position embedding and a transformer encoder module are utilized to train trajectory embeddings, allowing the model to learn sequential features and contextual information of trajectories. Moreover, a parallel recurrent neural network is employed to extract hidden trajectory features, capturing the differences between normal and anomalous trajectories. Finally, a linear layer is applied to fuse the spatiotemporal features and output the probability of an anomalous trajectory, enhancing the detection of vehicle trajectory anomalies. Experimental results on Beijing and Porto datasets demonstrate that the proposed PRNNT model significantly outperforms the iBAT (Isolation-Based Anomalous Trajectory), ATDC (Anomalous Trajectory Detection and Classification), ATD-RNN (Anomalous Trajectory Detection using Recurrent Neural Network), XGBoost (Extreme Gradient Boosting), GM-VSAE (Gaussian Mixture Variational Sequence AutoEncoder), and UA-OATD (Deep Unified Attention-based Sequence Modeling for Online Anomalous Trajectory Detection) models, achieving at least a 3.8%, 22.7%, 3.8%, 22.7%, 15%, and 16.7% improvement in F1-score, respectively.