Spatiotemporal GAN with Multi-Head Attention for Vehicle Trajectory Denoising

High-precision and high-quality vehicle trajectory data form the foundation for core applications in Intelligent Transportation Systems (ITS), such as traffic flow analysis, path planning, and autonomous driving. However, the trajectory data collected in practice often suffer from quality issues due to factors such as sensor noise, data transmission loss, and environmental interference, which seriously affect the performance of subsequent applications. To address this challenge, this paper proposes an innovative framework based on Generative Adversarial Networks (GANs), aimed at efficiently denoising and completing trajectory data. The framework deeply integrates a multi-head attention mechanism, a spatiotemporal convolutional network, and residual connections, and is optimized using a dual loss function. Specifically, the multi-head attention mechanism effectively captures long-range temporal dependencies within the trajectory data, significantly enhancing the model’s ability to understand complex motion patterns. The spatiotemporal convolutional network, through multi-scale feature extraction and residual connections, strengthens the model's perception and representation of local spatiotemporal features, and effectively mitigates gradient issues during deep network training. The combination of the dual loss function (reconstruction loss and adversarial loss) ensures that the generated data closely approximates the real data distribution while maintaining high-precision reconstruction. This study validates the proposed model using real UAV-collected vehicle trajectory data from the open traffic data platform of the Swiss Federal Institute of Technology Lausanne (EPFL). The dataset includes trajectory data from five continuous 30-minute time windows during the morning peak period on October 24, 2018 (08:30 − 11:00). Experimental results show that the proposed model demonstrates exceptional performance on the UAV trajectory dataset. Compared to traditional methods, trajectory reconstruction accuracy is significantly improved, with an average improvement rate of 85.6%, and the correlation coefficients of key features (such as latitude, longitude, and speed) exceed 0.92. This research provides an effective solution to improve the quality of trajectory data in intelligent transportation systems and holds significant theoretical and practical value for applications in autonomous driving, urban planning, and traffic management.