Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Online Residual Learning from Offline Experts for Pedestrian Tracking

Anastasios Vlachos, Anastasios Tsiamis, Aren Karapetyan, Efe C. Balta, John Lygeros

TL;DR

This paper proposes Online Residual Learning (ORL), a method that combines online adaptation with offline-trained predictions and treats the augmented lower-level predictors as experts, adopting the Prediction with Expert Advice framework.

Abstract

In this paper, we consider the problem of predicting unknown targets from data. We propose Online Residual Learning (ORL), a method that combines online adaptation with offline-trained predictions. At a lower level, we employ multiple offline predictions generated before or at the beginning of the prediction horizon. We augment every offline prediction by learning their respective residual error concerning the true target state online, using the recursive least squares algorithm. At a higher level, we treat the augmented lower-level predictors as experts, adopting the Prediction with Expert Advice framework. We utilize an adaptive softmax weighting scheme to form an aggregate prediction and provide guarantees for ORL in terms of regret. We employ ORL to boost performance in the setting of online pedestrian trajectory prediction. Based on data from the Stanford Drone Dataset, we show that ORL can demonstrate best-of-both-worlds performance.

Online Residual Learning from Offline Experts for Pedestrian Tracking

TL;DR

This paper proposes Online Residual Learning (ORL), a method that combines online adaptation with offline-trained predictions and treats the augmented lower-level predictors as experts, adopting the Prediction with Expert Advice framework.

Abstract

In this paper, we consider the problem of predicting unknown targets from data. We propose Online Residual Learning (ORL), a method that combines online adaptation with offline-trained predictions. At a lower level, we employ multiple offline predictions generated before or at the beginning of the prediction horizon. We augment every offline prediction by learning their respective residual error concerning the true target state online, using the recursive least squares algorithm. At a higher level, we treat the augmented lower-level predictors as experts, adopting the Prediction with Expert Advice framework. We utilize an adaptive softmax weighting scheme to form an aggregate prediction and provide guarantees for ORL in terms of regret. We employ ORL to boost performance in the setting of online pedestrian trajectory prediction. Based on data from the Stanford Drone Dataset, we show that ORL can demonstrate best-of-both-worlds performance.
Paper Structure (13 sections, 28 equations, 3 figures, 1 table, 3 algorithms)

This paper contains 13 sections, 28 equations, 3 figures, 1 table, 3 algorithms.

Table of Contents

  1. Introduction
  2. Problem Setting
  3. Online Residual Learning
  4. Residual Error Learning
  5. Combining Multiple Predictions
  6. Regret Guarantees
  7. Pedestrian Tracking
  8. Dataset
  9. Offline Model
  10. Simulation Results
  11. Conclusion and Future Work
  12. Proof of \ref{['thm:regret_guarantees']}
  13. Multiple-steps ahead prediction

Figures (3)

  • Figure 1: Schematic of the Online Residual Learning method. We attempt to learn the residual errors of the offline experts and combine them online.
  • Figure 2: Simulation results for the Hyang0 scene. Plots depicting the instantaneous and cumulative errors, showcasing how each prediction method performs over the horizon $T$, are depicted in (a), (b), respectively. In (c), the predicted trajectories, as well as the ground truth one are plotted on a sample photo of the scene. For visualization reasons, trajectories are downsampled by a factor of 30.
  • Figure 3: Simulation results for the Hyang3 scene. Plots depicting the instantaneous and cumulative errors, showcasing how each prediction method performs over the horizon $T$, are depicted in (a), (b), respectively. In (c), the predicted trajectories, as well as the ground truth one are plotted on a sample photo of the scene. For visualization reasons, trajectories are downsampled by a factor of 30.

Theorems & Definitions (1)

  • proof