Physical Distancing and Crowd Density Monitoring
( a ) ( b )
Fig . 2 : Camera calibration for distance calculation . a ) The original image with four reference markers drawn as the red polygon around the cashier stations . b ) The perspective transform applied relative to these markers rectifies the image , correcting for perspective distortion . This allows us to map pixel-distances to true physical distances .
Clustering
After we perform person detection on a per-frame by basis and person-tracking across frames , we can monitor the formation of groups and clusters and track them over time . Combined with the video calibration to convert pixel distances within the image to true physical distances in the physical space , we can detect when any number of people are involved in a physical distancing violation . In this project , we have used computer vision algorithms to perform person detection and tracking from video-input , but the methodology for monitoring groups and clusters only depends on the positions of each person over time . If other input sensors are used to detect positions of people in the space over time , the subsequent analytics for physical distancing still hold .
At each frame , we use the position of every detected person and compute all pairwise distances . These distances are true physical distances . We then apply single-linkage clustering 8 to connect any two people that are within a threshold , such as 6 feet or 2 meters . The visualization on Figure 3 shows a dot for each detected person as well as edges between any pair of people who are within 6 feet of each other . We also use color to show when people are close or dangerously
8 Gower JC , Ross GJ . Minimum spanning trees and single linkage cluster analysis . J R Stat Soc Ser C ( Applied Stat . 1969 ; 18 ( 1 ): 54-64 .
IIC Journal of Innovation - 7 -