Research On Behavior Prediction Of Ball Carrier In Basketball Video

As a typical team sports, basketball match has won great favor among the public with hundreds of millions of fans all over the world. This thesis made an in-depth study in the behavior prediction of ball carrier in basketball video, centering on the relevant problems:basketball court and court line extraction, player tracking, vision range determination of ball carrier etc. The main contribution can be concluded in the following aspects:(1) An improved K-means clustering is adopted to extract the basketball court for removing the interference of auditorium and narrowing the range of player tracking. Then the Hough transform and least square method are utilized to extract the court lines within the basketball court for the purpose of follow-up camera calibration.(2) An improved Particle Filter based on hierarchy feature fusion is proposed to track fast moving players in the basketball videos. More specifically, the player is represented by several auxiliary features and one main feature, which consist of a hierarchical structure and layers of the structure are linked through the proposal distribution of the sampling particles. The parameter update for each feature take places hierarchically so the simpler player features, which are updated first, guide the search in the state space of the more complex player features to relevant regions. This strategy enables the proposed algorithm to form better proposal distributions thus the parameter space of the complex feature is narrowed, the required number of particles is reduced and the search efficiency of particles is enhanced, leading to lower computational complexity. Additionally, a method to adapt the simpler features to cope with player occlusion is also proposed, in which the simpler features seeming to lose track based on a measure of their compatibility with the complex feature are deleted. When the number of the simpler features is low, new ones are added. In this way, the resistance to partial occlusions is increased.(3) A head pose recognition approach based on FSAMME is proposed for the ball carrier in low resolution video, which can be further used to determine his vision range. Aiming at the cluttered background, fast motion of the sportsmen and the low resolution of the head images in the basketball video, the covariance descriptor is adopted to efficiently fuse multiple visual feature of head region, fully exploiting both the self-correlation of these features and cross-correlation among them. According to the topological property that covariance descriptor can be formulated as differentiable manifold, it is mapped to the tangent space, in which head pose recognition is regarded as a multi-classification problem. The adopted multi-classifier is based on a new AdaBoost algorithm termed FSAMME(forward stagewise additive modeling using a multi-class exponential loss function), which only requires the performance of each weak classifier to be better than random guessing (rather than1/2) and directly complete the multi-class classification instead of reducing the multi-class classification problem into multiple two-class problems. As a result, the computational complexity is reduced and real-time performance is improved.(4) An approach based on online RBFNN is proposed to predict the ball carrier’s behavior——shooting, passing and dribbling in basketball matches. In order to describe the factors affecting the behavior of ball carrier, artificial potential field (APF)-based player information is introduced to model the court situation of all players after tracking and vision range determination, then a feature vector is formed as the input of the online RBF neural network. The behavior prediction of the ball carrier is solved by the online RBF neural network based on GIRAN learning algorithm. Compared with the offline RBF neural network, the online neural network can adjust both structure and parameters to basketball matches, thus the prediction accuracy is improved to some extent.