Description
Trajectory prediction is a necessary component of an autonomous driving pipeline because it is essential for safe and comfortable planning. However, most existing work in prediction only measures prediction performance in isolation on static datasets, as opposed to performance when used in conjunction with the rest of the autonomous driving pipeline. We show that commonly used prediction evaluation metrics on static datasets do not capture the full range of challenges faced by trajectory prediction algorithms in dynamic scenarios, and that no single prediction method works best for all situations. To do this, we implement and benchmark a variety of state-of-the-art prediction approaches, and evaluate their performance in conjunction with planning in simulated driving scenarios.