| Kinematic panorama is a wide-angle view that combines multiple kinematic movements taking place at different time phases into a static background image. Kinematic panorama can be of greater assistance for athletes and coaches to overview movement process and correct sport maneuver. In a common kinematic panorama system, background subtraction algorithm plays an important role, which separates foreground objects from background. Therefore, this thesis discusses a practical kinematic panorama system based largely on a novel background subtraction.First of all, prior arts of background subtraction are surveyed. The general model for background subtraction as well as several previous works are explored. Based on the above, an improved background subtraction based on Gaussian Mixture model is proposed, which follows a Bayesian paradigm to estimate model parameters and make decision on foreground or background. The key features of algorithm include modeling weights of Gaussian Mixture model by Dirichlet conjugate prior distribution, separating foreground from background by Online EM method, adopting optical flow information as high level feedback and a factor for foreground mixture. Experiments demonstrate that on the premise of comparable accuracy above-mentioned algorithm can be of less complexity and use less memory than prior arts.Also, architecture of an interactive kinematic panorama system is introduced. This semi-auto system can automatically make the rough approximation of foreground objects from image sequences acquired at key phases. After interactive refinement, more elaborate foreground objects are extracted and kinematic panorama is synthesized accordingly. In addition, details in function modules, data structures, interactive specifications and hardware equipments are listed. This system is used in field analysis for high jump and javelin programs.At the last part of this thesis, contributions of this work are summarized and further researches are also suggested. |
PDF Full Text Request