Automatic Surveillance Camera Calibration by Observation of Rigid Objects
This work is focused on automatic camera calibration based on multiple observations of arbitrary rigid objects. Based on observations of rigid objects moving in a common plane, we are able to calibrate camera w.r.t. the plane, and thus we are able to do measurements in a scene. Objects in the image plane are detected, and classified, and landmarks on these objects are localized. Our motivation was the usage of these methods in traffic scenarios, and thus as our ``objects'' we consider vehicles.
We propose three different methods that are able to compute camera calibration based on these localized landmarks in an image plane with the only limitation --- 3D models must be provided, but these can be known to the calibration system as a background. The camera calibration process is then fully automatic, and no more information is needed. Contrary to previous state-of-the-art methods for automatic camera calibration, the proposed methods are able to estimate all camera parameters (including focal length).We also collected a new dataset BrnoCarPark, which contains records of different scenes with detected vehicles and localized landmarks. Ground-truth measurements in scenes are available, and these can be re-computed by computed camera calibration parameters. All the proposed methods outperform the recent state-of-the-art method in an accurate manner. We evaluated our methods on the constructed dataset and also another dataset BrnoCompSpeed. We also made experiments on synthetic datasets, which prove the stability and usability of the proposed methods.
Automatic Camera Calibration, Rigid Objects, Calibration Dataset, Vehicle Detection, Vehicle Classification, Landmarks Localization, Vehicle Re-Identification, Horizon Estimation, AI City Challenge