Estimation of visual object trajectory by using video from multiple cameras
Sayed Masih Emami, Davide Moro · Chalmers Publication Library (Chalmers University of Technology) · 2012
Many effective techniques have been proposed on intelligent vision systems fortracking video objects using multiple cameras during the last decades.Since video from multiple cameras provides rich information from different view angles and locations, object tracking using video captured from multiple cameras is usually more robust for tracking objects with full/partial occlusions and intersections.One of the previously proposed methods uses homography for estimating 3D object trajectories on the ground plane, where images from multiple cameras and a synthetic top-view plane of the scene are used.The method is rather promising.This thesis focuses on estimating 3D object trajectories from videos captured by multiple cameras, where top-view images are not available, and cameras are not calibrated.Conventionally, estimating 3D object trajectories requires that either all cameras be calibrated, or Epipolar Geometry is used.One attractive way is to build 2D-3D relations without camera calibrations.In our study the trajectory of 3D object movement on the ground plane is derived by multiple uncalibrated cameras with the help of Epipolar Geometry without synthetic top-view scenes.This would allow cameras changing their positions or orientations during measurements.In our studies, correspondences among different camera view images from a same scene are established by using object point features from SIFT and RANSAC.This enables to establish relations between 2D-2D and 2D-3D images/object, hence, objects in 2D image planes and in the 3D world coordinate system.Within the framework of Epipolar Geometry, object positions on the ground plane are estimated up to a projective ambiguity in 3D-space (i.e.equality up to an arbitrary projective transformation in 3D-space) by minimizing re-projection error.The methods have been tested on videos captured from 3 IP cameras (recorded by ourselves).Two types of scenarios are tested to verify our algorithm for estimating time trajectories of 3D object in the ground plane.Visual inspections of the resulted trajectories are shown to be consistent to the real movement of object.