ONLINE LIVESTREAM CAMERA CALIBRATION FROM CROWD SCENE VIDEOS

Open Access
Author:
Bandyopadhyay, Anindita
Graduate Program:
Computer Science and Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
November 10, 2017
Committee Members:
  • Dr. Chita Das, Thesis Advisor
  • Dr. Robert Collins, Thesis Advisor
Keywords:
  • pedestrian detector
  • ground truth
Abstract:
The primary goal of this work is to analyze crowd scene videos to perform camera calibration using a data-driven method. Instead of using video files that are already available for performing computer vision research, the ones studied in this work have been recorded from publicly available online livestream video sources. Doing this will help to test the feasibility of calibrating remote webcams with minimal setup. Camera calibration is performed by studying the relationship between a person's foot location and height in varying spatial regions throughout the video frames. Obtaining information about a person's height as a function of foot location, "good" detections of a pedestrian detector (detections with reasonable heights) are filtered from the "bad" ones (detections with implausible heights). In the process, the camera's line of horizon is also computed which is related to the elevation angle of the camera. The key focus of this work includes generating video files of crowd scenes, filtering a pedestrian detector's results, and computing the camera's line of horizon.