Contrasting Estimation of Pattern Prototypes for Anomaly Detection in Urban Crowd Flow (ProtoDetect)

This repository is the official implementation of ProtoDetect.

Dataset

• All the dataset used in the paper are public available

  • NYC Open Data https://data.cityofnewyork.us/

  • Chicago Data Portal https://data.cityofchicago.org/

• A sample dataset is provided in this repository for test

  • dataset/NYC_dynamic_201410_minmax.npy the crowd flow dynamics of NYC in October 2014

  • dataset/NYC_graph_dict.json the dictionary of subgraphs and adjacency matrices

Test

1. Install requirements

2. Train

   # train the local and global ST-encoders
   python Extractor.py --mode local
   python Extractor.py --mode global
   
   # train ProtoDetect
   python ProtoDetect.py --mode train

   These scripts train the models on NYC dataset and save their weights in training_cache/ directory.

3. Test

   python ProtoDetect.py --mode eval

   This script outputs the flattened anomaly scores anomaly_score.npy.

Citation

@article{wang2024contrasting,
  title={Contrasting Estimation of Pattern Prototypes for Anomaly Detection in Urban Crowd Flow},
  author={Wang, Yupeng and Luo, Xiling and Zhou, Zequan},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2024},
  volume={},
  number={},
  pages={1-15},
  publisher={IEEE},
  keywords={Prototypes;Anomaly detection;Feature extraction;Self-supervised learning;Spatiotemporal phenomena;Behavioral sciences;Tensors;Anomaly detection;crowd management;urban transportation systems;contrastive learning},
  doi={10.1109/TITS.2024.3355143}
}

Acknowledgement

We appreciate the following github repos a lot for their valuable code:

• https://github.com/dleyan/STGAN

• https://github.com/RElbers/info-nce-pytorch

• https://github.com/uchidalab/time_series_augmentation