This repository contains the official implementation of C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-aware Trajectory Prediction (AAAI 2025).
You can find more information about the paper in the link below. C2F-TP
The NGSIM dataset contains trajectories of real freeway traffic captured at 10 Hz over a time span of 45 minutes in 2015. It is collected on eastbound I-80 in the San Francisco Bay area and southbound US 101 in Los Angeles. Like the baselines, the NGSIM dataset in our work is segmented in the same way as in the most widely used work Deo and Trivedi, 2018, so that comparisons can be made. Please download the NGSIM dataset from the NGSIM official website.
The highD dataset consists of trajectories of 110000 vehicles recorded at 25 Hz, which are collected at a segment of two-way roads around Cologne in Germany from 2017 to 2018. Due to the policy requirements of this dataset, please request and download the HighD dataset from the highD official website. Normally, applications take 7-14 working days to be approved.
We consider a two-stage training strategy to train C2F-TP as follows, which first trains the Refinement module and then trains the Spatio-Temporal Interaction module.
cd train
python train_denoise.py
python train_c2f.py
Run the following scripts to evaluate C2F-TP.
cd train
python evaluate_c2f.py
Create a new Python environment (c2f) using conda:
conda create -n c2f python=3.7
conda activate c2f
Run the following script for environment configuration.
pip install -r requirements.txt