Under the situations of big events, citywide crowd dynamics would manifest some different patterns from the normal routine. In this study, we propose a highly deployable system to extract the trend from momentary changes, and leverage multitask learning to predict crowd density and crowd flows simultaneously.
├── common # Parameters, data_utils
├── crowdFlowEncoder # CNN autoencoder for crowd flow
├── meshdynamic # meshgrid use
├── predCrowdDensity # baselines for predicting crowd density
│ ├── Historical Average
│ ├── Copy Last Frame
│ ├── ARIMA
│ ├── VAR
│ ├── ST-ResNet
│ ├── CNN / CNN enc-dec.
│ └── ConvLSTM / ConvLSTM enc-dec.
├── predCrowdEncodedFlow # baselines for predicting encoded crowd flows
│ ├── Historical Average
│ ├── Copy Last Frame
│ ├── ST-ResNet
│ ├── CNN / CNN enc-dec.
│ └── ConvLSTM / ConvLSTM enc-dec.
└── predMultitask(Ours)
├── Multitask CNN enc-dec.
└── Multitask ConvLSTM enc-dec. # our proposed model
Here are the working environment and major dependencies we run the code:
If you find anything in this repository useful to your research, please cite our paper :) We sincerely appreciate it. DeepUrbanEvent: A System for Predicting Citywide Crowd Dynamics at Big Events
@inproceedings{jiang2019deepurbanevent,
title={DeepUrbanEvent: A System for Predicting Citywide Crowd Dynamics at Big Events},
author={Jiang, Renhe and Song, Xuan and Huang, Dou and Song, Xiaoya and Xia, Tianqi and Cai, Zekun and Wang, Zhaonan and Kim, Kyoung-Sook and Shibasaki, Ryosuke},
booktitle={Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
pages={2114--2122},
year={2019},
organization={ACM}
}