This repository has code to calculate the number of social distancing violations given in an image. We implemented a CNN to identify ground plane and perform homography transformation to get bird's eye view of the image. We ran object detection algorithm - YOLOv3 to identify all the persons. If distance between bounding boxes in the bird's eye view is < 6 feet, it is considered a violation of social distancing.

Has source code and pretrained model of YOLOv3 algorithm used for object detection

• Code to build ground truth of ground plane using openCV
• CNN to predict homography matrix

Code to run object detection, warp images and calculate pairwise distances between people and count the number of social distancing violations(< 6 feet apart)

An existing implementation that uses similar approach of object detection and distance estimation to evaluating social distancing

We used Yang and Yurtsever's paper and code here for the people counting baseline. Please give them a reference if you used their work.

@misc{yang2020visionbased,
      title={A Vision-based Social Distancing and Critical Density Detection System for COVID-19}, 
      author={Dongfang Yang and Ekim Yurtsever and Vishnu Renganathan and Keith A. Redmill and Ümit Özgüner},
      year={2020},
      eprint={2007.03578},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}

To run the people counting with Faster R-CNN, open the Google Colab Notebook baseline.ipynb

Link to our dataset: In form of individual images and video for the baseline model (to run with Faster R-CNN)