26 January 2024
09:00 Master's Defense Room 53 of IC2
Theme
Graph-Based Spatio-Temporal Data Analysis for Passing Difficulty Classification in Football
Student
Maisa Silva
Advisor / Teacher
Ricardo da Silva Torres - Co-supervisor: Allan da Silva Pinto
Brief summary
Football, one of the most popular sports in the world, generates billions of dollars
annually. The recent availability of equipment for data collection and analysis
of football matches made it possible to apply mathematical models. These models
allow you to represent and understand how the dynamics of players and teams influence
events and match results. Several researches have applied mathematical models
to describe the patterns of football matches, evaluate their components and provide
crucial information for decision-making by technicians and teams, both to plan
strategies for matches and training procedures. One of the most common events
in a football match is the pass, which consists of transmitting the ball from a player
to another. Many metrics are studied in relation to passing, such as the frequency between
two players, the number of successful passes, the accuracy of the pass, among others. At the
However, the difficulty of the pass is still a topic little explored in scientific literature.
Our study presents an approach to characterize passes at different degrees of
Difficulty: easy, medium and difficult. We evaluate which representations are adopted by the elements
of the graph (vertices, edges and metrics) are more suitable for modeling the problem.
We then compare the impact of greater temporal detail on performance
pass classification. Finally, we evaluate whether the fusion of feature vectors
influences pass classification performance. The most prominent results include
the observation that graphs containing only the team that has possession of the ball are not a
good representation for the problem. Furthermore, the best graph representation for
binary temporal resolution presents edges for players up to 5 meters away,
with degree being the best measure of complex network. In this case, the balanced accuracy
reached 61%. For the multilevel evaluation, the best temporal resolution was 20 frames
with a bipartite graph configuration, in which the edges represent the interference
of the opponent in relation to the target, and the best measure was closeness; this set of vectors
of features achieved an accuracy of 65%, the same value as the agreement index
of the two labeling technicians. The combination of feature vectors, however,
did not present a significant gain in balanced accuracy, reaching a value of 66%.
These results suggest that the best feature vectors do not provide views
complementary to the classification task.
Examination Board
Headlines:
| Allan da Silva Pinto | LNLS/CNPEM |
| Milton Shoiti Misuta | FCA/UNICAMP |
| Bruno Luiz de Souza Bedo | EEFE/USP |
Substitutes:
| Hélio Pedrini | IC / UNICAMP |
| Tiago Guedes Russomanno | Technical University of Munich |