Core Concepts
G-PECNet improves upon the state-of-the-art PECNet model for pedestrian trajectory prediction through architectural improvements and synthetic data augmentation, achieving a 9.5% reduction in Final Displacement Error on the Stanford Drone Dataset benchmark.
Abstract
The paper introduces G-PECNet, an improved adaptation of the PECNet model for pedestrian trajectory prediction. The key contributions are:
- Augmenting the training dataset with synthetic trajectories generated using Reinforcement Learning (RL) and Hidden Markov Models (HMMs) to capture a wider range of pedestrian behaviors.
- Incorporating Sinusoidal Representation Networks (SIRENs) as the activation function to better capture high-frequency spatial and temporal details in the trajectories.
- Proposing a novel "Abruptness Score" metric to quantify the non-linearity of trajectories, which was used to guide the synthetic data generation process.
Experiments on the Stanford Drone Dataset (SDD) show that G-PECNet achieves state-of-the-art performance on the Final Displacement Error (FDE) metric, outperforming previous methods by 9.5%. The authors also provide detailed ablation studies on the effects of data augmentation and the decoupling of Average Displacement Error (ADE) and FDE.
Stats
The maximum Abruptness Score (AbScore) in the SDD dataset is 494866.37.
The minimum AbScore in the SDD dataset is 0.0.
The mean AbScore in the SDD dataset is 3430.665.
The standard deviation of AbScore in the SDD dataset is 11987.34.