Physical Distancing and Crowd Density Monitoring
tracking-by-detection approach 55 which is based on tracking people detections across frames using the intersection over union between detections in consecutive frames . Thus , the person detection results for each frame are provided as inputs to the tracking model and the tracker outputs a trajectory of movement for each person . The tracker provides a unique ID associated with each person ’ s trajectory . These IDs are randomly generated by the tracking model and are not associated with the identity of the person such employee name or employee ID . Further , to ensure that the identity information is not compromised , our solution blurs the pixels in the bounding box of person detection before visualization . Thus , our solution detects and tracks people anonymously since it does not have any information about the identity of the people .
Distance Calculation via Camera Calibration
The deep learning model outputs the coordinates of a detected person relative to the input image and not relative to the physical space . To calculate distances between pairs of people we must first account for radial and tangential optical distortion 6 due to the camera and apply a transformation to the coordinates . In this example from our SAS café camera , the input image is taken as a crop from an area towards the center of the field-of-view of the camera and we observe no radial distortion . However , the ground is not parallel to the imaging plane , so we must correct for perspective distortion causing objects further away to shrink . We correct for this using a perspective transform 7 aligned with some reference markers in our physical space ( See Figure 2 ). After applying the transform , we can approximate true physical distances between every pair of points .
5 Bochinski E , Eiselein V , Sikora T . High-Speed tracking-by-detection without using image information . 2017 14th IEEE Int Conf Adv Video Signal Based Surveillance , AVSS 2017 . 2017 ;( August ). doi : 10.1109 / AVSS . 2017.8078516 .
6
Wolf L , Shashua A . On Projection Matrices Pk→P2 , k = 3 ,..., 6 , and their Applications in Computer Vision . Int J Comput Vis . 2002 ; 48:53-67 .
7 Haralick RM . Using perspective transformations in scene analysis . Comput Graph Image Process . 1980 ; 13 ( 3 ): 191- 221 .
- 6 - November 2020