DenseTNT: End-to-end Trajectory Prediction from Dense Goal Sets

M2I: From Factored Marginal Trajectory Prediction to Interactive Prediction

“…More recently, a family of models are proposed to improve prediction accuracy and coverage by first predicting high-level intentions, such as goal targets [11,13,15,26,30,42], lanes to follow [20,33], and maneuver actions [8,9], before predicting low-level trajectories conditioning on each intention. Such models demonstrate great success in predicting accurate trajectories for single agents in popular trajectory prediction benchmarks, such as Argoverse [6] and Waymo Open Motion Dataset [10].…”

“…Such models demonstrate great success in predicting accurate trajectories for single agents in popular trajectory prediction benchmarks, such as Argoverse [6] and Waymo Open Motion Dataset [10]. While our proposed approach M2I can use an arbitrary prediction model, we choose to adopt an anchor-free goal-based predictor [15] because of its outstanding performance.…”

“…Rasterized representation draws traffic context on a single image with multiple channels and excels at capturing geometrical information. Both representations have achieved top performance in trajectory prediction benchmarks such as Argoverse and WOMD [6,10,12,14,15].…”

“…Recent advances have shown great success in predicting accurate trajectories by learning from real-world driving examples. Many existing trajectory prediction works [5,12,15,23,25,34] focus on generating marginal prediction samples of future trajectories over individual agents, failing to reason about their interactions in the future. As a result, the prediction samples over multiple agents may overlap with each other and result in sub-optimal performance.…”

“…To generate scene compliant trajectories, one can learn a joint predictor to predict trajectories in a joint space over multiple agents; however, it suffers from an exponentially increasing prediction space as the number of agents increases. Taking the popular goal-conditioned model [15] as an example, while it is feasible to predict a set of goals for a single agent, the goal space increases exponentially with the number of agents and becomes unmanageable for even two agents with a few hundred goal candidates for each agent. A more computationally efficient alternative to producing scene compliant trajectories is to post-process marginal prediction samples by pruning colliding ones; however, such an ad-hoc approach fails to take into account potential agent interactions in the future and may ignore other conflicts which are hard to prune by heuristics.…”