|
||||
Abstract of master's workAuthor: Sereda Andrey Theme: "The automated system of video recognition for sports competitions" AbstractMy work is devoted to development of the automated system of video analysis system for football coaching. A purpose of this system is to analyze video records of a football match and on their basis count and visualize various data. The system consists of 2 parts: subsystem of video analysis and subsystem of trajectories analysis. Entrance data of video analysis subsystem are:
Target data of analysis subsystem are trajectories of objects on a field during entire match. The subsystem of trajectories analysis calculates on obtained trajectories various statistical data, models the separate moments of a match by animation, etc. In my work the attention is given first of all to development of a video analysis subsystem. Expected deadline - January, 2007. The given abstract fixes a current state of development. The purposes of this work are:
As is known, football is extremely popular game. Football clubs are ready to pay greater money for any development which can raise efficiency of work of the trainer. Therefore the development of that system is demanded in the market. Similar systems already exist, however they possess the certain lacks:
Apparently, development of system, which is capable to use a small number of chambers (2-4), which can be established not in strictly certain places, and also as much as possible automating process of the video analysis of and minimizing manual input is actual. Used algorithmsThe problem of object tracking in video records can be broken into 2 subtasks:
Background subtraction is used for image segmentation. Background image corresponds to a view of field without players and ball. This image can be obtained automatically from sequence of frames with objects. To allocate objects in the frame it is enough to compare pixel-by-pixel this frame to a background image. Pixels for which a difference of color less than threshold value, are defined as concerning to a background, and for what it is more - as concerning to objects. This method is not capable to separate the objects which is partially overlays each other. So, in final version of system more accurate segmentation algorithm will be used. Model of a player movement is used to predict area in which it can be in next moment of time. In each time after prediction if this areas objects are searched in corresponding areas of frames. After finding location of all players next time moment will be analyzed. When segmentation algorithm can't recognize some situation, decision holds over the end of analysis. Then operator must solves all unsolved situation manually. Current results
Test of current realized algorithm was done. 2 video records with resolution 320x200 and 20 frames was created by means of 3D modeling. Each record corresponds one camera. There are 4 moving objects and 2 collisions in this scene. Trajectories of objects have been found without mistakes. On a computer with processor Athlon XP 2200 + the analysis takes 0.6 sec. On real video record with resolution 720x580 and 60 frames with 13 players the analysis takes 3.2 sec. However, quality of recognition at the given stage of development has appeared insufficient for practical use. Now I have work on improvement of quality of object tracking on real video records with poor quality. |
||||
|