CloudFast-DL4PuDe: A Cloud-based Deep Learning Framework for Early Detection of Pushing at Crowded Event Entrances

This repository is for:

1. The submitted paper:

Alia, Ahmed, Mohammed Maree,  Mohcine Chraibi, Anas Toma, and Armin Seyfried. "A Cloud-based Deep Learning Framework for Early Detection of Pushing at Crowded Event Entrances." 2023.

2. The preprint version:

Alia, Ahmed, et al. "A cloud-based deep learning system for improving crowd safety at event entrances." arXiv preprint arXiv:2302.08237 (2023).

Goal

The framework aims to early detect pushing patches automatically in running and crowded event entrances, particularly in the live camera stream of entries.

Motivation

The motivation of the framework is to help organizers and security forces to intervene early and mitigate dangerous situations.

Table of Contents

• The Architicture of CloudFast-DL4PuDe framework
• How to use CloudFast-DL4PuDe framework
• Demo
• Building and training the CNN architectures
• Trained CNN models
• Dataset
• Video experiments
• Computational time evaluation

The Architicture of CloudFast-DL4PuDe Framework

How to Use CloudFast-DL4PuDe Framework

1. Click on

2. File--> save a copy in drive.

3. Download the cloud-fast-system directory.

4. Upload the cloud-fast-system directory to Google Colab at the same location of the copy of CFSystem.ipynp.

    --Colab NoteBooks
         --copy of CFSystem.ipynp
         --cloud-fast-system
   

5. For the input stream, we recommend using a virtual camera with one of the video recordings of crowded event entrances. Video experiments are available in the video directory. Kinly note that OBS studio is free and open source software for virtual camera and live streaming.

6. Run copy of CFSystem.ipynp

Demo

Client side

1. How to run
   • Launch the virtual camera based on the entrance2 video experiment and OBS studio software.
   • Open the Jupyter Notebook (copy of CFSystem.ipynp) and entering the required inputs.
     • patch=[2,4]
     • roi=[107,55,669,270]
2. Output example: video of the live camera stream with annotations of the predicted pushing patches, red boxes represent the predicted pushing patches.

CNN Models

Building and Training the CNN Architectures

1. Clone the repository in your directory.

git clone https://github.com/abualia4/CloudFast-DL4PuDe.git

2. Install the required libraries. TensorFlow is the main required library

pip install tensorflow

3. Open this notebook and follow the guides.

Trained CNN Models

The trained CNN models are available in the links below.

• Trained models in related work.
• Trained popular CNN models.

Dataset

The dataset generated and used during this work are available from the corresponding authors upon request.

Video Experiments

The original video experiments that were used in this work are available through the Pedestrian Dynamics Data Archive hosted by the Forschungszentrum Juelich. Moreover, the undistorted videos are available through this link.

Computational Time Evaluation

The results of the computational time evaluation of our and the baseline systems for each video experiment are available here.

Acknowledgement

1. Thanks to the authors of RAFT repository and the corresponding paper.
2. Thanks to the author of Access Webcam for Images and Video notebook .