carrotsniper/multiclass-sgcn

Multiclass-SGCN: Sparse Graph-based Trajectory Prediction with Agent Class Embedding

This is the official implementation of our paper Multiclass-SGCN https://arxiv.org/abs/2206.15275.

This work has been accepted at the International Conference on Image Processing (ICIP 2022). The conference paper is available here.

Abstract

Trajectory prediction of road users in real-world scenarios is challenging because their movement patterns are stochastic and complex. Previous pedestrian-oriented works have been successful in modelling the complex interactions among pedestrians, but fail in predicting trajectories when other types of road users are involved (e.g., cars, cyclists, etc.), because they ignore user types. Although a few recent works construct densely connected graphs with user label information, they suffer from superfluous spatial interactions and temporal dependencies. To address these issues, we propose Multiclass-SGCN, a sparse graph convolution network based approach for multi-class trajectory prediction that takes into consideration velocity and agent label information and uses a novel interaction mask to adaptively decide the spatial and temporal connections of agents based on their interaction scores. The proposed approach significantly outperformed state-of-the-art approaches on the Stanford Drone Dataset (SDD) , providing more realistic and plausible trajectory predictions.

Models

All the experiments are done with Stanford Drone Dataset (SDD), we compare our proposed work with 8 models.

Method	mADE	mFDE
Linear	37.11	63.51
SF	36.48	58.14
Social-LSTM	31.19	56.97
Social-GAN	27.25	41.44
CAR-Net	25.72	51.80
DESIRE	19.25	34.05
Social-STGCNN	26.46	42.71
Semantics-STGCNN	18.12	29.70
Multiclass-SGCN (ours)	14.36	25.99

Installation

The codebases are built on top of Semantic-STGCNN and SGCN

Requirements

• Linux or Windows with Python 3.9, with CUDA verion: 11.3
• PyTorch ≥ 1.5 and torchvision that matches the PyTorch installation.
• ALl libraries: install from requirements.txt

  # For conda user
  conda create --name <env_name> --file requirements.txt
  
  # For pip user
  pip install -r requirements.txt
  
  

Steps

1. Download the codes:

https://github.com/Carrotsniper/Multiclass-SGCN.git
cd Multiclass-SGCN

2. To train the model run

python train.py

3. To test the model run

python test_pred.py

Cite this Work

If this work is useful, please consider citing the paper, and/or mentioning this repository:

@inproceedings{li22multiclasssgcn,
 author={Li, Ruochen and Katsigiannis, Stamos and Shum, Hubert P. H.},
 booktitle={Proceedings of the 2022 IEEE International Conference on Image Processing},
 series={ICIP '22},
 title={Multiclass-SGCN: Sparse Graph-based Trajectory Prediction with Agent Class Embedding},
 year={2022},
 publisher={IEEE},
 location={Bordeaux, France},
}

multiclass-sgcn's Issues

Source of txt files of dataset

Hello! I'm working on trajectory visualization. I see that the content of the txt file in the dataset you are using is different from the official SDD data set. Where do these files come from? Because the frame id in the dataset you provided is inconsistent with the official one. I want to learn how the files are preprocessed.