1Depa men o Compu e Science, Uni e si y o Pue o Rico a Río Pied as
2Facul ad de Ciencias en la Salud, Uni e sidad Manuela Bel án, Bogo á, Colombia
3Depa men o Physical Educa ion and Rec ea ion, Uni e si y o Pue o Rico a Río Pied as
In oduc ion
A Low-Cos App oach o Spo s Pe o mance Assessmen
Using Compu e Vision and A i icial In elligence Tools
You Only Look Once: Yolo
2
Re e ences
The sys em de eloped akes spo s aining ideos and uns he Yolo 11 algo i hm o e hem,
p omp ing he use o highligh key a eas o he ideo (such as goals and ball s a ing posi ions)
which i la e uses alongside he Yolo model's ou pu o analyze he aining and e alua e playe
pe o mance. Re u ns he a emp 's di ec ion, accu acy, as well as he playe 's eac ion ime.
The p og am also ou pu s he analyzed ideo (displaying ball ajec o y and size which has o be
p edic ed when no de ec ed by Yolo) and a ious g aphs/da a o u he analysis, (See Figu e
3). Cu en ly we use Google Colab [5] o un he p og am, as i o e s GPUs be e sui ed o
compu a ionally in ensi e asks.
Resul s
The aining en i onmen ha inspi ed his p ojec ,
de eloped by Ja ie Oso io and based on s udies
by Musculus e al. [1] and Knöbel & Lau enbach [2],
plays an impo an ole in he unc ioning o he
so wa e. We a e de eloping an ecologically alid
en i onmen , ha is, one ha in eg a es
pe cep ual, cogni i e, and mo o skills wi hin
speci ic spo s con ex s [3]. The sys em consis s o
a p ojec o o gene a e a isual s imuli o he
playe and a sma phone o cap u e ideos
con aining he a hle e’s eac ions. Figu e 1 shows a
schema ic ep esen a ion o he aining
en i onmen , in which he playe is p esen ed wi h
wo possible socce ield goals. Figu e 2 illus a es
an example o a playe ’s kick a e he ideo has
been p ocessed using YOLO, wi h ball- acking
added du ing pos -p ocessing.
T adi ional spo s aining o en elies on manual
obse a ions and subjec i e eedback, which can
be slow, inconsis en , and p one o e o . Exis ing
spo s aining sys ems and specialized so wa e
end o be expensi e and di icul o implemen ,
making hem inaccessible o many uni e si y
spo s eams. In his wo k, we explo e how a i icial
in elligence and compu e ision can be applied o
ack, analyze, and e alua e a hle ic pe o mance
mo e objec i ely and e icien ly wi hin a low-cos
aining amewo k.
T aining En i onmen
1. Musculus, L., Lau enbach, F., Knöbel, S., Reinha d, M. L., Weigel, P., Ga zmaga, N., ... & Pelka, M. (2022). An
assis o cogni i e diagnos ics in socce : wo alid asks measu ing inhibi ion and cogni i e lexibili y in a
socce -speci ic se ing wi h a socce -speci ic mo o esponse. F on ie s in psychology,13, 867849.
2. Knöbel, S., & Lau enbach, F. (2023). An assis o cogni i e diagnos ics in socce (Pa II): De elopmen and
alida ion o a ask o measu e wo king memo y in a socce -speci ic se ing. F on ie s in psychology,13,
1026017.
3. Renshaw, I., Da ids, K., A aújo, D., Lucas, A., Robe s, W. M., Newcombe, D. J., & F anks, B. (2019). E alua ing
weaknesses o “pe cep ual-cogni i e aining” and “b ain aining” me hods in spo : An ecological dynamics
c i ique. F on ie s in psychology,9, 2468.
4. Ul aly ics. (n.d.). YOLO h ead-sa e in e ence. Ul aly ics YOLO Docs. Re ie ed Augus 19, 2025, om
h ps://docs.ul aly ics.com/guides/yolo- h ead-sa e-in e ence/
5. Google. (n.d.). Google Colabo a o y. Re ie ed Augus 19, 2025, om h ps://colab. esea ch.google.com
Ou sys em uses YOLO o "You Only Look Once," Ul aly ics' eal ime objec de ec ion deep
lea ning model, which h ough i s obus algo i hm is able o quickly and accu a ely iden i y
di e en objec s in images [4].
Yolo 8. We o iginally s a ed wi h some demos using Yolo 8, a e sion o he model ha is able
o p ocess images e y quickly bu sac i ices some accu acy.
Yolo 11. We swi ched o Yolo 11 a e some ime due o Yolo 8's issues wi h de ec ing objec s
consis en ly and accu a ely a longe anges. We decided o use Yolo 11l, a e sion o Yolo 11
which has a ound 25.3 million pa ame e s (making he p og am slowe , bu much mo e
accu a e).
2
Acknowledgemen s
This wo k was pa ially suppo ed by he Na u al Sciences Dean’s O ice a UPR
Río Pied as,D . Al a ez’s Seed Funds, and D . O ozco’s FADI Funds. Special
hanks o Ja ie Oso io o gene ously sha ing ideas om his doc o al disse a ion.
Gab iel To es1Ca los Vázquez1Ja ie Oso io2,3 Edusmildo O ozco1Michael Al a ez1,*
gab iel. o es54@up .edu ca los. azquez78@up .edu ja ie .oso io2@up .edu
edusmildo.o ozco1@up .edu
michael.al a ez2@up .edu
LATIN AMERICA HIGH PERFORMANCE COMPUTING CONFERENCE
Sep embe 22-26
Kings on, Jamaica
Conclusions
This sys em uses he YOLO 11 objec de ec ion model o analyze spo s aining ideos by
au oma ically iden i ying key elemen s like goal a eas and ball posi ions. I compu es me ics
such as sho accu acy, eac ion ime, and di ec ion, and ou pu s anno a ed ideos wi h
p edic i e ball ajec o ies alongside pe o mance g aphs. Running on Google Colab wi h GPU
accele a ion, he implemen a ion suppo s pa allel p ocessing h ough Py hon
mul ip occesing, o e ing signi ican speed gains. Designed o ealis ic game scena ios,
a o dable cos , and a simple implemen a ion, he ool shows p omising po en ial o enhance
coaching eedback and suppo playe de elopmen in uni e si y-le el spo s
p og ams. Addi ionally, he in eg a ion o gami ica ion s a egies such as poin sys ems,
challenges, and in e ac i e eedback could u he inc ease playe engagemen , mo i a ion,
and lea ning ou comes, making he ool e en mo e e ec i e in aining en i onmen s.
Figu e 1. Expe imen al se up: p ojec ion-based en i onmen
displaying wo possible socce ields o pe cep ual and
decision-making aining.
Figu e 2. Example o he ou pu ideo p ocessed wi h YOLO
and ajec o y highligh ed in pos -p ocessing.
Figu e 3. T ajec o ies ob ained a e pos -p ocessing YOLO labels o quan i y
he playe ’s accu acy. The do ed line indica es he poin o he sho s, and he
socce ields a e included as pa o he playe ’s isual con ex .
