Decrease Longitude Diversity in Lane-conditioned Trajectory Prediction

Predicting the future trajectories of surrounding agents is crucial for achieving safe autonomous driving. In addition to the strong prior information provided by maps, our key insight is that trajectories of surrounding vehicles can also guide trajectory prediction models. This leads us to develop a road supplement framework, a solution that integrates scene information in a deterministic manner during the preprocessing stage. Subsequently, we evaluates the probability of driving on the supplemented and reachable lane. Furthermore, we focus on achieving the highest possible prediction accuracy with the fewest prediction outputs. To accomplish this, we employ a decoder based on Conditional Variational Autoencoder (CVAE) that outputs the most probable set of trajectories to the goal lane. Detailed ablation experiments demonstrate the effectiveness of decreasing longitudinal diversity using CVAE. We benchmark our model on the nuScenes dataset and achieve favorable performance on metrics with fewer outputs.