social-distancing-monitoring

A Vision-based Social Distancing and Critical Density Detection System for COVID-19

Developed by Dongfang Yang and Ekim Yurtsever at Control and Intelligent Transportation Research (CITR) Lab, The Ohio State University.

Paper: arXiv preprint

Updata Log:

• 2020.09.15: We have added the instructions for running the pedestrian detection and the data analysis. Please see the section of 'Getting Started'.
• 2020.07.10: We are still in the process of finalizing the repository. The complete version will be released soon!

System Overview

Our system is real-time and does not record data. An audio-visual cue is emitted each time an individual breach of social distancing is detected. We also make a novel contribution by defining a critical social density value for measuring overcrowding. Entrance into the region-of-interest can be modulated with this value.

Social Distancing Monitoring

Illustration of pedestrian detection and social distancing monitoring.

Camera calibration

NYC Grand Central terminal: We found the floor plan of the building and calibrated the camera by picking landmarks. We provide the transformation matrix in calibration/grand_central_matrix_cam2world.txt

Oxford town center: The original dataset provides the transformation matrix. We added it here also calibration/oxford_town_matrix_cam2world.txt

Mall: We could not found the transformation matrix or the floor plan of this dataset. Instead, we first estimated the size of a reference object in the image by comparing it with the width of detected pedestrians and then utilized the key points of the reference object to calculate the perspective transformation. We provide this transformation matrix in calibration/mall_matrix_cam2world.txt

Critical Density

Keeping the social density under the critical value will keep the probability of social distancing violation occurrence near zero with the linear regression assumption.

Getting Started

1. Environment Config

The program was developed based on python 3.7 with pytorch 1.5. We highly recommend to use conda. After you have created a new conda environment, use the following command to install pytorch:

conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch

You may also need to install the packages specified in requirements.txt file if you don't have them.

2. Download Dataset

We provide an alternative link for you to download all the datasets: https://drive.google.com/file/d/1G6nZS-EZLrNBC68CRDf-yo2uj3k32j35/view?usp=sharing

Then copy all the files into folder datasets in the repository.

3. Run Pedestrian Detection

Execute detect.py to obtain the detection result. The result will be saved at folder results in the repository. Result will be saved as pickle .p file for each dataset.

4. Run Analysis

Execute analyze.py to obtain the analysis result. It will be saved in the same folder results.

TODO Lists

• Social distancing monitoring pipeline
• Evaluation on different pedestrian crowd datasets
• Detector: Faster R-CNN
• Detector: Yolo v4
• Detector: EfficientDet
• Critical density analysis
• Embedded system integration
• Camera calibration UI

Contact

For further info please contact:

• Dongfang Yang: [email protected]
• Ekim Yurtsever: [email protected]