Tracking of U.S. Olympic Boxers Justin Brewer April 28, 2014 Project Goals • Segment the boxers and the referee and determine there position on the boxing ring. – Origin taken to be bottom left corner of the ring. • Produce X & Y plots vs. time for each target. • Produce overlaid and individual heat maps for each person. • Aim for a position accuracy of 20cm. What does the data look like. X: 145 565-145 = 420 pix/23 ft. => 0.5991pix/ft. X: 565 20 cm = 12pixels Approach • User input captures initial position of each corner of the ring, boxers and referee. • Use the motion of the boxers against a stationary background to aid in segmentation. • Apply Morphological Operators • In trouble locations use mean-shift or CAMshift along with previous boxer location . • Create requested plots from position data. “Basic” Background Subtraction Improved Adaptive Gaussian Mixture Model for Background Subtraction • The value of a pixel at time t in RGB space denoted by • Pixel-based background subtraction involves decision if pixel belongs to background (BG) or some foreground (FG) object by the Bayesian decision R: • For the general case we assume nothing is known about the foreground. • Decide if a pixel belongs to the foreground if: where is a threshold value. Z.Zivkovic, Improved adaptive Gausian mixture model for background subtraction, International Conference Pattern Recognition, UK, August, 2004. Improved Adaptive Gaussian Mixture Model for Background Subtraction • is the background model. • The background model is modeled by a Gaussian mixture model (GMM) with components estimated by a training set of images denoted as • Training set is chosen for a reasonable set of images and is continually updated adding and discarding new images from the video. • The GMM is modeled by: Covariance Matrix Variance Mean Mixing Weights Shadow Detection • Using the mean and the variance vector. Chrominance of a pixel the difference between expected color of a pixel and value in current image. • Pixel is then classified. is Segmentation Results Tracking Referee (Proof of concept) Heat Map (Proof of concept) What to finish this week • Implement CAMshift or Mean-shift and expand tracking to account for merging blobs. • Quantify accuracy. Acknowledgments Thank you to the U.S. Olympic committee for supplying the boxing videos and project goals. Thank you Dr. Hoff for your support on this project so far. Questions