Real-time tracking of sports pitch markings

Short presentation of G.Thomas's paper

Transcript

1. Real-Time Camera Tracking using Sports Pitch Markings G. Thomas (BBC) Journal of Real Time Image Processing, Vol. 2, No. 2-3, November 2007, pp. 117-132
2. Paper’s Goal • Placing virtual annotations on the pitch Distances, whether a player is off-side, …
3. What do we need • For this to happen we need: – Camera position, orientation & focal length – In real time – Using image analysis (pitch markings) metric camera calibration
4. Novel Ideas of this paper • Computing camera position – Is improved using multiple images. • Initializing the tracking process – An automatic method is presented. Using ideas from Hough transform.
5. Paper Assumptions • We know the pitch model • Camera is PTZ • Tracking is real-time • Tracking initialization takes ~1s • “Sufficient Accuracy” – Which is not good for 3D reconstruction.
6. Sufficient Accuracy
8. Camera Position • Fixed during a match • Needs to be estimated once • We have camera position for Azadi dataset.
9. Camera Position - Problems • Estimating camera position and focal length simultaneously is hard. • Use multiple images with wide range of pan & tilt. • Use a single optimization for all images. – Different pan, tilt and zoom. – Same position & pitch rotation
10. Initialization - Approach • Hough space – Angle of the line – Shortest distance to the center of the image • Each line in the image is a bin (point) in Hough space
11. Initialization - Approach • If the camera pose is known (approx) It is possible to find peaks in Hough space that correspond to each pitch line.
12. Initialization - Approach • Important feature about this paper: • Lets’ not use Hough to find lines and then find correspondence
13. Initialization - Approach • Let’s say we estimate a pose. • How good is this estimated pose? 1. we know the pitch model (set of 3D lines) 2. project those lines to image using the pose (set of 2D lines) 3. find those lines in Hough space (set of Hough bins) 4. match-value = sum of found Hough bins
15. Initialization - Approach The best pose has the highest match-value? Which is fast
16. Initialization - Details • Pre-processing – Chroma keyer – Line detector
17. Implementation https://github.com/yassersouri/thomas-calibration