Pedestrian Trajectory Forecasting Using Deep Ensembles Under Sensing Uncertainty

This paper introduces a novel approach for addressing a fundamental challenge in pedestrian trajectory prediction: reliable future state estimation in the presence of sensor noise and uncertain perception. It is common for traditional prediction models to produce overconfident and error-prone deterministic estimates. Previous research exploring probabilistic estimation methods often overlooked inherent noise in upstream perception data which is a crucial factor under adverse conditions like bad weather or occlusion. While Bayes filters are adept at integrating data from noisy sensors, they falter in handling non-linearities and long-term forecasting. Our proposed approach is an end-to-end estimator capable of handling noisy sensor data to deliver robust predictions of future states with uncertainty bounds while factoring in upstream perception uncertainty. This is achieved through an innovative encoder-decoder-based deep ensemble network designed to capture both perception and prediction uncertainty during trajectory prediction. We benchmark our model against other established approximate Bayesian inference methods on publicly available pedestrian datasets. The results demonstrate that our deep ensembles not only yield more robust predictions but also capable of reliable predictions on out-of-distribution pedestrian trajectories. Our method plays a key role in improving dynamic agent trajectory prediction, offering a more precise and robust framework for future state estimation.