Research On Key Issues In Soccer Video Analysis

The number of digital video clips has grown explosively, which leads to an emerg-ing challenge on how to use these videos effectively. In order to access specific contentin massive video media, it is urgent to index the video data by its semantic content. Inresponse to this demand, video content analysis technologies emerged and had attract-ed attention from areas of both academia and industry. Soccer is quite popular, whichmakes soccer video become one of the most valuable video types that need to be ana-lyzed. Over the past decade, some achievements have been gained in the soccer videoanalysis. However, there is still room for improvement, especially in summarizing andsolving the specific circumstances of broadcast soccer video analysis. Therefore, we aremotivated to design new algorithms according to these specific circumstances.To solve the existing problems in the previous research, theoretical basis and devel-opment status of the soccer video analysis techniques are analyzed in detail. Then, severalkey problems demonstrated in the soccer video content analysis process are figured out.With the guidance of human visual cognition principle and scenario restriction of soccervideo content analysis, key issues of soccer video analysis are studied, which broadensthe research approach of soccer video analysis. According to this research thinking, sev-eral key technologies in soccer video analysis are addressed in this disseratation, whichinclude playfield detection, player detection, ball tracking and attack tactic analysis. Themain research contents and contributions of this dissertation are as follows:(1) A playfield detection method exploiting both color and local consistency featureis proposed. Color feature are used in existing playfield detection, which is not effectiveto remove the green pixels that not belong to playfield. To solve this problem, localconsistency feature is introduced, and the playfield is detected using both color feature andlocal consistency feature. To determine the detection threshold of local consistency, a twodimensional histogram based method and a color constrained Otsu(c Otsu) based methodare proposed, which are based on the principle of color characteristic and local entropycharacteristic of playfield pixels respectively. Experiments show that the proposed methodis more effective and is able to detect playfield in several typical environments.(2) An automatic player detection method based on fuzzy decision making one-classSVM is proposed. Detection results of statistical classifier player detection methods arebetter than rule based player detection methods. However, the manually labelled train-ing samples are used in these statistical classifier based player detection methods. Thus,the cost is very high. To resolve this problem, an automatic player detection method us-ing fuzzy decision making one-class SVM and automatically collected player samplesare proposed. In this method, one-class SVM(OCSVM) is introduced to train playerdetector by drawing lessons from the human object category classification mechanism.What’s more, decision function of OCSVM is improved by dividing the decision valuedynamically using the fuzzy decision method, which is able to reduce the detection er-ror caused by the insu?cient representativeness of the automatically collected trainingsamples. Finally, a set of criterion is introduced to obtain the training samples automat-ically, and player detection experiments are performed on these training samples usingFD-OCSVM. Experiments show that better detection results are obtained using the pro-posed method in the scenario of using automatically collected training samples, whichimprove the automatic degree of player detection.(3) A ball detection method based on class weighted spatial Fuzzy C-means(ws-FCM) and a ball tracking method based on multiple search regions dynamic kalman filter(MDKF) are proposed. In the aspect of ball detection, as the size of the ball is too small toextract distinguishable feature, it is di?cult to detect the ball automatically. To solve thisproblem, an automatically ball detection method is proposed. In this method, the targetfunction of the spatial Fuzzy C-means is improved first. Then, a bi-threshold strategy isproposed to detect the ball automatically. Experiments show that the ball can be detectedautomatically using the proposed method. In the aspect of ball tracking, existing meth-ods will lost the ball when it is occluded by several players successively. To solve thisproblem, motion state of the soccer ball in broadcast soccer video is analyzed, which isinspired by the contextual cueing effect of human visual search. According to the motionstate of the soccer ball, the parameters updating function of dynamic kalman filter(DKF)is improved. Thus, multiple search regions dynamic kalman filter(MDKF) is proposed,which enhances the robustness of soccer ball tracking by extending search area. Exper-iments show that the proposed method can track the ball more robust, which with betterocclusion handle ability.(4) An attack tactic analysis method using ball deviation ratio is proposed. Play re-gion sequence is used as the feature of attack pattern analysis in existing methods, whichwill cause the loss of ball position information. To solve this problem, a novel ball po-sition feature using relative position information between soccer ball and symbol linesis proposed in light of the intrinsic reference theory of human spatial perception, whichis named as ball deviation ratio. Then, the problem of how to extract ball deviation ra-tio from broadcast soccer videos is further discussed. To achieve this goal, an automaticoffside line detection method by combing directional filter, mean-biteral filter and houghtransform is proposed. Using the extracted ball deviation ratio, attack tactic patterns areanalyzed. Experiments show that the proposed method can detect the tactic more accu-rately, and the sub-patterns can also be analyzed using the proposed method.