Identifying Basketball Plays from Sensor Data; towards a Low-Cost Automatic Extraction of Advanced Statistics
Author
Arbues Sanguesa, Adria
Term
4. term
Publication year
2017
Submitted on
2017-06-08
Pages
92
Abstract
Avancerede statistikker er et vigtigt redskab for basketballtrænere til at forbedre træning og taktik. Ved at analysere, hvordan spillere agerer under bestemte forhold, kan holdets præstation optimeres yderligere. I USA leverer firmaer som STATS og Second Spectrum sådanne statistikker til alle NBA-hold via avancerede opsætninger med flere kameraer, men prisen ligger langt uden for budgettet for de fleste europæiske hold. Som et mere overkommeligt alternativ præsenteres her en første prototype baseret på positioneringssensorer, der registrerer spillernes placering. Der blev opbygget et eksperimentelt datasæt, og centrale kendetegn ved basketballspil blev udtrukket. Med Support Vector Machines (en maskinlæringsmetode) blev der opnået 97,9% nøjagtighed i at identificere fem klassiske spilmønstre: “floppy offense”, “pick and roll”, “press break”, “post-up” og “fast break”. Når disse spilmønstre kan genkendes i videosekvenser, kan avancerede statistikker derefter udledes langt lettere.
Advanced statistics are an important tool for basketball coaches to improve training and tactics. By analyzing how players behave under specific conditions, a team’s performance can be further optimized. In the United States, companies such as STATS and Second Spectrum provide these statistics to all NBA teams using complex multi-camera systems, but the cost is far beyond the budgets of most European teams. As a more affordable alternative, this work presents a first prototype based on positioning sensors that capture players’ locations. An experimental dataset was created, and key basketball features were extracted. Using Support Vector Machines (a machine learning method), 97.9% accuracy was achieved in identifying five classic plays: “floppy offense,” “pick and roll,” “press break,” “post-up,” and “fast break.” Once these plays can be recognized in video sequences, advanced statistics can then be generated much more easily.
[This abstract was generated with the help of AI]
Keywords
Documents