Основні поняття
A novel Bayesian framework is proposed that explicitly models the relationship between homographies of consecutive video frames, as well as the uncertainty in keypoint measurements, to significantly improve existing methods for soccer field registration.
Анотація
The proposed approach, Bayesian Homography Inference from Tracked Keypoints (BHITK), employs a two-stage Kalman filter framework to estimate the homography between a video frame and a soccer field template.
The first stage is a linear Kalman filter that tracks the image keypoint positions, considering the estimated affine transformation between consecutive frames. The second stage is an extended Kalman filter that incorporates the homography as part of the state vector, explicitly modeling the relationship between homographies of consecutive frames as well as the uncertainty in the field template keypoint positions and the measured image keypoints.
BHITK can be easily integrated with existing keypoint detection methods. It enables less sophisticated and less computationally expensive keypoint detection approaches to outperform state-of-the-art methods in most homography evaluation metrics.
Furthermore, the authors have refined the homography annotations of the existing WorldCup and TS-WorldCup datasets and released the consolidated and refined WorldCup (CARWC) dataset for public use.
Статистика
"The mean matrix entries over all j of ΣI,j, estimated as described in 4.1, are [4.95, -0.06; -0.06, 0.95]."
"The mean entries of the estimated covariance matrix of the homography obtained with RANSAC (used to initialise the homography elements of the Kalman filter state vector) are [254, 1.76, 0.06, 123, 22.81, -0.03, -28795, 947, 1.76, 0.25, 0.00, 0.54, 0.05, 0.00, -163, -1.17, 0.06, 0.00, 0.00, 0.03, 0.01, 0.00, -6.51, 0.21, 123, 0.54, 0.03, 60.23, 11.22, -0.02, -14023, 472, 22.81, 0.05, 0.01, 11.22, 2.18, 0.00, -2605, 87.42, -0.03, 0.00, 0.00, -0.02, 0.00, 0.00, 3.85, -0.12, -28795, -163, -6.51, -14023, -2605, 3.85, 3280363, -108856, 947, -1.17, 0.21, 472, 87.42, -0.12, -108856, 4186]."
Цитати
"The proposed method, Bayesian Homography Inference from Tracked Keypoints (BHITK), employs a two-stage Kalman filter and significantly improves existing methods."
"BHITK can be easily integrated with existing keypoint detection methods. It enables less sophisticated and less computationally expensive methods to outperform the state-of-the-art approaches in most homography evaluation metrics."