Built a framework which takes as input a video and performs social distancing detection. This framework detects humans within the video, applies homography techniques to adjust for depth and distance, detects whether humans are at least a safe distance of each other (2 meters), and outputs a labelled video in which social distancing violations are visually marked. This framework implements and allows two leading object detection approaches to be used interchangeably (Faster R-CNNs and YOLOv3). Information is logged as it runs, thus allowing for comparisons between the two approaches.

We use this framework to detect social distancing in 4 different video datasets with both Faster R-CNNs and YOLOv3, and perform analysis on the data. We derive several insights on social distance detection from this data and conclude with a comparison on the performance of Faster R-CNNs versus YOLOv3 for social distancing detection, and we finally present that a social distancing detector using Faster R-CNNs correctly detects social distancing violations 2.52 times more frequently than YOLOv3.

Learn more about our Social Distancing Detector in our project report: https://github.com/khizr/SocialDistancingDetector/blob/master/Social%20Distancing%20Detector%20Report.pdf

Labelled Output

Google drive containing RCNN inputs and results:* https://drive.google.com/drive/folders/1xJaAs5CqOp1v7XIeuWg6aY2MvpPEFRP0?usp=sharing

Google drive containing YOLOv3 inputs, weights file, and results:* https://drive.google.com/drive/folders/1AdVufg4cprv0mZhkhSBLyFnoJU6aM5rH?usp=sharing

Install dependencies: (use cu101 because colab has CUDA 10.1) !pip install cython pyyaml==5.1

Install detectron2 !pip install detectron2==0.1.3 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.5/index.html

Note: if the above doesn't work, try pip3 install torch torchvision Note: install a CUDA version of torch, see https://pytorch.org

Install darknetpy (Object detection with YOLO) https://pypi.org/project/darknetpy/

Then download weights file for "yolov3-608" from https://pjreddie.com/darknet/yolo/ or download the file from our YOLOv3 Google Drive (labelled yolov3.weights) and place it in the same directory as detector.py. (The file is 250MB and is too large to directly submit)

Get info on script: ./detector.py -h

Example script run command (RCNN using CPU): ./detector.py sample.mp4 RCNN -1 logs.json true