Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

G-PECNet: Towards a Generalizable Pedestrian Trajectory Prediction System

Aryan Garg, Renu M. Rameshan

TL;DR

Problem: robust generalization of pedestrian trajectory prediction for autonomous drones navigating dynamic environments. Approach: G-PECNet extends PECNet with RL-based synthetic trajectory augmentation using HMM interactions, SIREN-based periodic activations for high-frequency spatial-temporal detail, and AbScore-based non-linearity analysis for data curation. Key findings: achieves state-of-the-art $FDE$ on the Stanford Drone Dataset benchmarks and demonstrates significant gains when training with augmented data; the 6% augmentation with SIREN yields notable improvements. Significance: enables more reliable downstream planning for social robots and drones, provides a geometry-inspired non-linearity metric for outlier detection, and shares the codebase for reproducibility.

Abstract

Navigating dynamic physical environments without obstructing or damaging human assets is of quintessential importance for social robots. In this work, we solve autonomous drone navigation's sub-problem of predicting out-of-domain human and agent trajectories using a deep generative model. Our method: General-PECNet or G-PECNet observes an improvement of 9.5\% on the Final Displacement Error (FDE) on 2020's benchmark: PECNet through a combination of architectural improvements inspired by periodic activation functions and synthetic trajectory (data) augmentations using Hidden Markov Models (HMMs) and Reinforcement Learning (RL). Additionally, we propose a simple geometry-inspired metric for trajectory non-linearity and outlier detection, helpful for the task. Code available at https://github.com/Aryan-Garg/PECNet-Pedestrian-Trajectory-Prediction.git

G-PECNet: Towards a Generalizable Pedestrian Trajectory Prediction System

TL;DR

Problem: robust generalization of pedestrian trajectory prediction for autonomous drones navigating dynamic environments. Approach: G-PECNet extends PECNet with RL-based synthetic trajectory augmentation using HMM interactions, SIREN-based periodic activations for high-frequency spatial-temporal detail, and AbScore-based non-linearity analysis for data curation. Key findings: achieves state-of-the-art on the Stanford Drone Dataset benchmarks and demonstrates significant gains when training with augmented data; the 6% augmentation with SIREN yields notable improvements. Significance: enables more reliable downstream planning for social robots and drones, provides a geometry-inspired non-linearity metric for outlier detection, and shares the codebase for reproducibility.

Abstract

Navigating dynamic physical environments without obstructing or damaging human assets is of quintessential importance for social robots. In this work, we solve autonomous drone navigation's sub-problem of predicting out-of-domain human and agent trajectories using a deep generative model. Our method: General-PECNet or G-PECNet observes an improvement of 9.5\% on the Final Displacement Error (FDE) on 2020's benchmark: PECNet through a combination of architectural improvements inspired by periodic activation functions and synthetic trajectory (data) augmentations using Hidden Markov Models (HMMs) and Reinforcement Learning (RL). Additionally, we propose a simple geometry-inspired metric for trajectory non-linearity and outlier detection, helpful for the task. Code available at https://github.com/Aryan-Garg/PECNet-Pedestrian-Trajectory-Prediction.git
Paper Structure (16 sections, 3 equations, 5 figures, 7 tables)

This paper contains 16 sections, 3 equations, 5 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Augmenting with RL Synthetic Trajectories
  4. Periodic Activation: SIREN Improvement
  5. Novel Loss and Evaluation Metric: AbScore
  6. Appendix
  7. Quantitative Comparison
  8. Stanford Drone Dataset: Data Analysis
  9. Deep Policy Gradient Network
  10. Architecture Brief
  11. Further Experiments
  12. Ablation: Decoupled ADE & FDE or No Social Pooling in PECNet
  13. Effects of Data Augmentations
  14. Baseline Metrics' Reproduction (PECNet)
  15. Non-linearity analysis: Abruptness-Score
  16. ...and 1 more sections

Figures (5)

  • Figure 1: The RL agent is inserted and trained in an HMM interaction playground. Agent's trajectory is turquoise. Evolution of the samples produced. First, the agent learns to turn. The second depicts a complicated scene where the agent learns to avoid multiple collisions. The last scene depicts the agent successfully avoiding a collision and reaching its goal.
  • Figure 2: RL modeling
  • Figure 3: SDD training dataset. Frobenius norm-based clustering of trajectories, with black trajectories representing the cluster.
  • Figure 4: Defining Turns
  • Figure 5: SDD's Non-Linearity Distribution Tail. (Thresholded to remove outliers)