Relative Camera Localisation in Non-Overlapping Camera Networks Using Multiple Trajectories

Abstract

In this article we present an automatic camera calibration algorithm using multiple trajectories in a multiple camera network with non-overlapping field-of-views (FOV). Visible trajectories within a camera FOV are assumed to be measured with respect to the camera local co-ordinate system. Calibration is performed by aligning each camera local co-ordinate system with a pre-defined global co-ordinate system using three steps. Firstly, extrinsic pair-wise calibration parameters are estimated using particle swarm optimisation and Kalman filtering. The resulting pair-wise calibration estimates are used to generate an initial estimate of network calibration parameters, which are corrected to account for accumulation errors using particle swarm optimisation-based local search. Finally, a Bayesian framework with Metropolis algorithm is adopted and the posterior distribution over the network calibration parameters are estimated. We validate our algorithm using studio and synthetic datasets and compare our approach with existing state-of-the-art algorithms.