Automatic camera calibration by landmarks on rigid objects

This article presents a new method for automatic calibration of surveillance cameras. We are dealing with traffic surveillance, and therefore, the camera is calibrated by observing vehicles; however, other rigid objects can be used instead. The proposed method is using keypoints or landmarks automatically detected on the observed objects by a convolutional neural network. By using fine-grained recognition of the vehicles (calibration objects), and by knowing the 3D positions of the landmarks for the (very limited) set of known objects, the extracted keypoints are used for calibration of the camera, resulting in internal (focal length) and external (rotation, translation) parameters and scene scale of the surveillance camera. We collected a dataset in two parking lots and equipped it with a calibration ground truth by measuring multiple distances in the ground plane. This dataset seems to be more accurate than the existing comparable data (GT calibration error reduced from 4.62 % to 0.99 %). Also, the experiments show that our method overcomes the best existing alternative in terms of accuracy (error reduced from 6.56 % to $$4.03\,\%$$ ) and our solution is also more flexible in terms of viewpoint change and other.