Hierarchical Latent Structure for Multi-modal Vehicle Trajectory Forecasting

Variational autoencoder (VAE) has widely been utilized for modeling data distributions because it is theoretically elegant, easy to train, and nice manifold representations. However, when applied image reconstruction synthesis tasks, VAE shows the limitation that generated sample tends be blurry. We observe a similar problem, in which trajectory located between adjacent lanes, often arises VAE-based forecasting models. To mitigate this we introduce hierarchical latent structure into model. Based on assumption distribution can approximated as mixture of simple (or modes), low-level variable employed model each mode high-level represent weights modes. accurately, condition using two lane-level context vectors computed novel ways, one corresponds vehicle-lane interaction other vehicle-vehicle interaction. The are also used via proposed selection network. evaluate our model, use large-scale real-world datasets. Experimental results show not only capable generating clear multi-modal but outperforms state-of-the-art (SOTA) models terms prediction accuracy. Our code available at https://github.com/d1024choi/HLSTrajForecast .