%0 Journal Article
%T Soccer Ball Detection by Comparing Different Feature Extraction Methodologies
%A Pier Luigi Mazzeo
%A Marco Leo
%A Paolo Spagnolo
%A Massimiliano Nitti
%J Advances in Artificial Intelligence
%D 2012
%I Hindawi Publishing Corporation
%R 10.1155/2012/512159
%X This paper presents a comparison of different feature extraction methods for automatically recognizing soccer ball patterns through a probabilistic analysis. It contributes to investigate different well-known feature extraction approaches applied in a soccer environment, in order to measure robustness accuracy and detection performances. This work, evaluating different methodologies, permits to select the one which achieves best performances in terms of detection rate and CPU processing time. The effectiveness of the different methodologies is demonstrated by a huge number of experiments on real ball examples under challenging conditions. 1. Introduction Automatic sport video analysis has become one of the most attractive research fields in the areas of computer vision and multimedia technologies [1]. This has led to opportunities to develop applications dealing with the analysis of different sports such as tennis, golf, American football, baseball, basketball, and hockey. However, due to its worldwide viewership and tremendous commercial value, there has been an explosive growth in the research area of soccer video analysis [2, 3] and a wide spectrum of possible applications have been considered [4¨C6]. Some applications (automatic highlight identification, video annotation and browsing, content-based video compression, automatic summarization of play, customized advertisement insertion) require only the extraction of low-level visual features (dominant color, camera motion, image texture, playing field line orientation, change of camera view, text recognition) and for this reason they have reached some maturity [7, 8]. In other applications (such as verification of referee decision, tactics analysis, and player and team statistic evaluations), instead, more complex evaluations and high-level domain analysis are required. Unfortunately, the currently available methods have not yet reached a satisfactory accuracy level and then new solutions have to be investigated [9]. In particular, the detection and localization of the ball in each frame is an issue that still requires more investigation. The ball is invariably the focus of attention during the game, but, unfortunately, its automatic detection and localization in images is challenging as a great number of problems have to be managed: occlusions, shadowing, presence of very similar objects near the field lines and regions of player¡¯s bodies, appearance modifications (e.g., when the ball is inside the goal, it is faded by the net and it also experiences a significant amount of deformation during
%U http://www.hindawi.com/journals/aai/2012/512159/