A Diffusion-TGAN Framework for Spatio-Temporal Speed Imputation and Trajectory Reconstruction

Generative Adversarial Networks (GAN) have been widely used in traffic data imputation to improve the accuracy of data imputation. However, existing GAN-based models often suffer from mode collapse and cannot fully reflect the complex characteristics of real-world traffic, which affects the quality of data imputation. To address these challenges, we incorporate the Diffusion Model (DM) into the GAN framework, integrating the traffic dynamics modeling process within the Diffusion-GAN network. Based on this, we propose a Diffusion-TGAN speed data imputation model to generate individual vehicle speeds. Combined with the generated vehicle speed, the group trajectory reconstruction result is further given. The model uses the forward process of DM to generate condition vectors to guide the training of GAN generator. Subsequently, the discriminator of GAN takes the traffic dynamics constraints into account during adversarial training. Traffic dynamics modeling aims to make the generated speed data consistent with the real traffic characteristics. Experiments on multiple data sets show that the proposed model effectively imputes in the spatio-temporal speed data, and reduces the RMSE of the speed considering the position by 23.4% compared with the common GAN model, and reduces the RMSE by 39.7% in the trajectory reconstruction respectively. The code and our model are available at GitHub.