Computer Vision Social Distancing Final Project
https://github.com/endernac/CV-Final-Project
There are two main components to this project:
This is intended to be run locally, since Google Colaboratory does not support XWindows. This notebook uses the OpenCV imshow with a few modifications to allow for simple homography selection. The user should select 4 points in the image which represent a square on the ground plane. The order of selection should be:
- bottom left
- bottom right
- top right
- top left
Double clicking on a point in the image will create a dot, indicating the selected location. Once four points have been selected, the user can double click near points to shift them. Additionally, a green grid will be displayed to help guide selection. This grid may be difficult to work with at first, as poor selections will make strange vanishing points. However, after adjustments are made, an ideal selection will make it very clear where the ground plane is. When the desired grid has been selected, hit ESC to return. The homography of interest will be displyed in the jupyter notebook, which can be copied (with comas added) into the Colaboratory notebook.
The following notebooks are meant to be run on Google Colaboratory. Two notebooks are provided, preloaded with information for a simple run of some VIRAT example data. people_detector.ipynb draws circles on an overhead view without any indication of violating social distance while people_detector_distancing_visualization.ipynb changes circle color based on violations. These notebooks are meant to be run on Google Colaboratory to leverage the GPUs provided.
These run with a basic pipeline:
- Homography loading (from corresponding VIRAT file, or manualy input after selection from computer)
- Detection loop (with calibration during the first frame with people)
Currently, the camera pose calculation is set up such that the angle under 90 degrees should be chosen.