paul-pias / social-distance-monitoring

Used Yolact++ to implement social distance monitoring.

Social-Distance-Monitoring

Introduction

This repository holds the implementation of monitoring social distancing implied for COVID-19 using YOLACT: Real-time Instance Segmentation) for object detection.

User Guideline

System Requirement

• For better performance you will cuda version 10 or 10.1
• Python3

Installation

    $ pip install -r requirements.txt

For the installation of torch using "pip" kindly follow the instructions from Pytorch

First, you need to clone the repository using gitbash (if gitbash is already installed) or you can download the zip file.

    $ git clone https://github.com/paul-pias/Social-Distance-Monitoring.git

If want to see your output in your browser execute the "server.py" script or else run "inference.py" to execute it locally.

If you want to run the inference on a ip camera need to use WebcamVideoStream with the following command.

    "rtsp://assigned_name_of_the_camera:assigned_password@camer_ip/"

If you want to use YOLACT++, compile deformable convolutional layers (from DCNv2). Make sure you have the latest CUDA toolkit installed from NVidia's Website.

    cd external/DCNv2
    python setup.py build develop

In the official Yolact repository there are several pre-trained model available.

Image Size	Model File (-m)	Config (-c)
550	yolact_resnet50_54_800000.pth	yolact_resnet50
550	yolact_darknet53_54_800000.pth	yolact_darknet53
550	yolact_base_54_800000.pth	yolact_base
700	yolact_im700_54_800000.pth	yolact_im700
550	yolact_plus_resnet50_54_800000.pth	yolact_plus_resnet50
550	yolact_plus_base_54_800000.pth	yolact_plus_base

Things to consider

Download the pre-trained weights and save in the folder weights, then from your terminal run the following command based on your preference.

    python inference.py -m=weights/yolact_base_54_800000.pth -c=yolact_base -i 0

Here 0 as id passed if you want to run the inference on webcam feed. If you don't parse any argument it will run with the default values. You can tweak the following values according to your preference.

Input	Value	Description
width, height	1280 x 720
display_lincomb	`False
crop	True	For better segmentation use this flag as True
score_threshold	0.15	Higher the value better the performace
top_k	30	At max how many objects will the model consider to detect in a given frame
display_masks	True	Draw segmentation
display_fps	False
display_text	True
display_bboxes	`True
display_scores	False
fast_nms	True
cross_class_nms	True
display_text	True

The Purpose!

Social Distancing is a way of protecting yourself and others around you by knowing the facts and taking appropriate precautions. To prevent the spread of COVID-19 few guidelines were provided by World Health Organzation (WHO) and other public health agency. Maintaining at least 1 metre (3 feet) distance between yourself and anyone who is coughing or sneezing or for the time being everyone was one of them.

To measure distance between two person eucledian distance was used in this work. Euclidean distance or Euclidean metric is the "ordinary" straight-line distance between two points in Euclidean space.

The Euclidean distance between two points p and q is the length of the line segment connecting them . In the Euclidean plane, if p = (p1, p2) and q = (q1, q2) then the distance is given by

This formula was applied in the draw_distance(boxes) function where we got all the bounding boxes of person class in a given frame from the model where each bounding is a regression value consisting (x,y,w,h) . Where x and y represent 2 co-ordinates of the person and w & h represent widh and height correspondingly. All combinations of boxes were found to calculate the distance between them.

Acknowlegement

Thanks to Daniel Bolya et. el for introducing Single Shot detection (SSD) implementation for segmentation in YOLACT & YOLACT++ as it becomes less memory hungry.

However, for more details on Object detection and Image understanding you can also visit this repo .

social-distance-monitoring's Issues

How is the distance threshold determined?

Hello,

In the code, I noticed that 250 is used as a threshold to classify people as being too close ("red"). However, this is a distance in pixels, not meters. How was it determined then?

Also, isn't that ignoring the problem of perspective? (if two people are close to the camera then their pixel distance will be large, but if they are far away from the camera than their pixel distance will be small).

Thanks,
David