Informed Spectral Normalized Gaussian Processes for Trajectory Prediction

TL;DR

This work proposes a novel regularization-based continual learning method for SNGPs, which enables the use of informative priors that represent prior knowledge learned from previous tasks, and applies it to the trajectory prediction problem in autonomous driving by integrating prior drivability knowledge.

Abstract

Prior parameter distributions provide an elegant way to represent prior expert and world knowledge for informed learning. Previous work has shown that using such informative priors to regularize probabilistic deep learning (DL) models increases their performance and data-efficiency. However, commonly used sampling-based approximations for probabilistic DL models can be computationally expensive, requiring multiple inference passes and longer training times. Promising alternatives are compute-efficient last layer kernel approximations like spectral normalized Gaussian processes (SNGPs). We propose a novel regularization-based continual learning method for SNGPs, which enables the use of informative priors that represent prior knowledge learned from previous tasks. Our proposal builds upon well-established methods and requires no rehearsal memory or parameter expansion. We apply our informed SNGP model to the trajectory prediction problem in autonomous driving by integrating prior drivability knowledge. On two public datasets, we investigate its performance under diminishing training data and across locations, and thereby demonstrate an increase in data-efficiency and robustness to location-transfers over non-informed and informed baselines.

Figures (4)

• Figure 1: The informed CoverNet-SNGP model consists of a spectral normalized feature extractor and a last layer Gaussian Process with a Fourier feature approximated radial basis function kernel. Given a Birds-Eye-View RGB rendering and the target's current state, the model classifies a set of candidate trajectories according to their drivability in task $i$ and their likely realization in task $i+1$. Our method regularizes the training on task $i+1$, given the MAP estimates and Laplace approximated covariance from task $i$ as informative priors, thereby integrating the drivability knowledge following the PIL approach.
• Figure 2: Average performance and standard deviation of 5 independent repetitions over decreasing subsamples of NuScenes (bold as best).
• Figure 3: Average performance and standard deviation of the informed and non-informed CoverNet-SNGP on Boston and Singapore test data, with (a) models trained on Singapore training data and (b) models trained on Boston training data (five repetitions).
• Figure 4: Average performance and standard deviation of 5 independent repetitions trained on Singapore and Boston locations from NuScenes.