Gait-Based Prediction of Penalty Kick Direction in Soccer

David Freire-Obreg

, Oliverio J. Santana

, Javier Lorenzo-Navarro

Daniel Hern

andez-Sosa

and Modesto Castrill

on-Santana

SIANI, Universidad de Las Palmas de Gran Canaria, Las Palmas de Gran Canaria, Spain

Keywords:

Penalty Kick Prediction, Soccer Analytics, Gait Analysis, LSTM, Action Anticipation, Sports Analytics.

Abstract:

Understanding and predicting penalty kick outcomes is critical in performance analysis and strategic decision-

making in soccer. This study investigates the potential of gait-based biometrics to classify the intended shoot

zone of penalty takers using temporal gait embeddings extracted from multiple state-of-the-art gait recognition

backbones. We compile a comprehensive evaluation across several models and datasets, including baseline

models and other models such as GaitPart, GLN, GaitSet, and GaitGL trained on OUMVLP, CASIA-B, and

GREW. A standardized LSTM-based classiﬁer is trained to predict the shooting zone from video-level gait

sequences, using consistent train-test splits to ensure fair comparisons. While performance varies across

model-dataset pairs, we observe that certain combinations yield better predictive accuracy, suggesting that

the gait representation and the training data inﬂuence downstream task performance to some degree. This

work demonstrates the feasibility of using gait as a predictive cue in sports analytics. It offers a structured

benchmark for evaluating gait embeddings in the context of penalty shoot zone prediction.

1 INTRODUCTION

Penalty shootouts represent one of professional soc-

cer’s most critical and psychologically demanding

scenarios. Despite their brief duration, these isolated

events often carry disproportionate weight in deter-

mining the outcome of tightly contested matches. In

recent FIFA World Cups, over a quarter of knockout-

stage games have been decided from the penalty

mark, highlighting this phase’s strategic and emo-

tional signiﬁcance. As a result, understanding the dy-

namics of penalty kicks has become increasingly rel-

evant for players, coaching staff, and analysts.

Given the high pressure and game-deciding na-

ture of penalties, the ability to anticipate the direc-

tion of a penalty shot could provide a strategic advan-

tage to goalkeepers and analysts alike. While physi-

cal attributes, such as body orientation and approach

angle, have been studied extensively, the subtler pre-

shot movement patterns—speciﬁcally, the shooter’s

gait—remain underutilized in predictive modeling.

Developing systems that can interpret a player’s

https://orcid.org/0000-0003-2378-4277

https://orcid.org/0000-0001-7511-5783

https://orcid.org/0000-0002-2834-2067

https://orcid.org/0000-0003-3022-7698

https://orcid.org/0000-0002-8673-2725

movement sequence before the shot and reliably pre-

dict shot direction could augment goalkeeper training

and real-time match analytics.

Penalty kick prediction and related tasks have

been approached by analyzing body motion, pose es-

timation, and temporal movement patterns, frequently

supported by player tracking and activity recognition

frameworks. For instance, random forests combined

with context-conditioned motion models to detect and

track players under dynamic conditions, enabling ac-

tion inference (Liu and Carr, 2014). A comprehen-

sive review of player tracking methods has outlined

the challenges posed by occlusion and pose variabil-

ity, pertinent to pre-kick movement analysis, such

as gait (Manaﬁfard et al., 2016). Interaction mod-

eling between players and the ball has also been in-

vestigated, incorporating physical constraints to im-

prove the interpretability of motion dynamics (Maksai

et al., 2016). Recently, penalty scenarios have been

addressed using Human Action Recognition (HAR)

models (Freire-Obreg

on et al., 2025). However, un-

like that work, which focuses solely on the running

and kicking stages by cropping the sequence, our ap-

proach considers the entire penalty sequence without

truncation at any moment. This decision is grounded

in the observation that, although gait models are ef-

fective at capturing individual walking patterns, their

142

Freire-Obregón, D., Santana, O. J., Lorenzo-Navarro, J., Hernández-Sosa, D. and Castrillón-Santana, M.

Gait-Based Prediction of Penalty Kick Direction in Soccer.

DOI: 10.5220/0013669900003988

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2025), pages 142-149

ISBN: 978-989-758-771-9; ISSN: 2184-3201

ultimate goal is subject identiﬁcation. Consequently,

we aim to provide as many temporal and contextual

cues as possible by preserving the full sequence. De-

spite these developments, the temporal modeling of

ﬁne-grained kinematic features, particularly gait dur-

ing a penalty approach, has received limited attention.

This gap highlights the potential for sequence-based

representations to enhance predictive modeling in this

context, mainly when used with recurrent neural ar-

chitectures.

In this work, we propose a deep learning approach

based on Long Short-Term Memory (LSTM) net-

works to predict the shot zone of a penalty taker using

their gait sequence leading up to the shot. We con-

struct a consistent experimental framework by lever-

aging several gait backbones pretrained with publicly

available datasets for penalty kick scenarios. Each

video is segmented into ﬁxed-length sequences, and

a neural network is trained to classify the ﬁnal shot

direction into one of three zones. To ensure fair com-

parisons, we apply a shared train/test split across all

datasets and repeat experiments multiple times, sav-

ing only the best-performing model for each.

Our contributions are threefold: (1) we design a

standardized and reproducible LSTM-based pipeline

for predicting penalty kick direction leveraging gait

pretrained backbones; (2) we benchmark this pipeline

across multiple gait datasets using consistent data

splits and repeated trials to account for variance in

training outcomes; and (3) we conduct a qualitative

error analysis of the best-performing models, reveal-

ing that most misclassiﬁcations occur between adja-

cent shoot zones. This indicates that gait patterns

associated with neighboring shot directions are often

subtly different and challenging to distinguish based

on pre-kick motion alone.

2 RELATED WORK

Computer vision applications in sports have evolved

to support coaching, broadcasting, and analytics

through player tracking, event recognition, and mo-

tion analysis. In soccer, this has led to technolo-

gies like TRACAB and Hawk-Eye, which are used

for player tracking and goal-line detection, respec-

tively (ChyronHego, 2017; Innovations, 2017). These

systems rely on calibrated multi-camera setups and

computer vision pipelines for real-time data extrac-

tion. However, their focus is primarily on posi-

tional and event-level data, offering limited insight

into biomechanical features such as gait.

Tracking players for tactical and performance

analysis is a signiﬁcant research focus. For instance,

Manaﬁfard et al. provide a survey of player track-

ing techniques in soccer, noting challenges such as

occlusions, appearance similarity, and erratic move-

ments (Manaﬁfard et al., 2016). Techniques ranging

from model-based detection to context-conditioned

motion models have been proposed to tackle these

problems (Liu and Carr, 2014). Still, these ap-

proaches focus on position rather than motion style,

leaving gait-speciﬁc analysis underexplored.

Motion analysis of athletes has traditionally re-

lied on marker-based motion capture systems, which

are impractical for in-game scenarios. Markerless

systems using multiple or single cameras have been

explored to visualize motion trails or generate pose

sequences (Figueroa et al., 2006). These visualiza-

tions are often used for coaching or broadcast en-

hancements but lack integration into predictive mod-

els. Furthermore, pose estimation accuracy under re-

alistic conditions, such as during soccer penalty kicks,

remains a technical challenge.

Event detection in sports has been another key

area of study. Kapela et al. proposed methods for de-

tecting goals, shots, and fouls through visual analysis

and scoreboard interpretation (Kapela et al., 2014).

Recent work has also examined the use of visual

and temporal features for predicting shot outcomes,

including the classiﬁcation of ball-on-goal positions

based on the kicker’s shooting action (Artiles et al.,

2024), as well as large-scale performance analyses

in other sports domains such as running, highlight-

ing the value of motion-based modeling across dis-

ciplines (Freire-Obreg

on et al., 2022). While this

supports high-level game understanding, such meth-

ods do not typically incorporate pre-shot motion cues,

such as approach gait, which may reveal a player’s in-

tention during set-pieces like penalty kicks.

Our work complements and extends this body

of research by applying gait embeddings, commonly

used for biometric identiﬁcation, to predict shot

zones in soccer penalty scenarios. Unlike previous

approaches focused on spatial position or detected

events, we explore how temporal motion patterns can

serve as predictive features. This represents a novel

intersection of gait recognition and sports analytics,

leveraging insights from both domains.

3 METHODOLOGY

This section details the overall approach used to

model and classify penalty kick directions based on

visual motion cues. Our methodology is structured

into three main components: formal problem deﬁni-

tion, gait-based feature extraction, and shot direction

Gait-Based Prediction of Penalty Kick Direction in Soccer

143

Gait

Backbone

Classifier

LSTM

Dropout

Dense

R C L

Figure 1: Pipeline overview. The kicker is isolated from each video to generate silhouette sequences encoded using a pre-

trained gait model. The resulting embeddings are processed by an LSTM-based classiﬁer to predict shot direction, R, C, and

L correspond to Right, Center, and Left from the goalkeeper’s perspective.

classiﬁcation. By using recent advances in human

motion analysis and lightweight temporal modeling,

we aim to evaluate the effectiveness of gait represen-

tations in anticipating a penalty kicker’s intent. Each

component is described in the subsections that follow.

3.1 Problem Deﬁnition

Let there be m instances of penalty shooters, where

each instance is deﬁned as a tuple o

(i)

= (V

(i)

, z

(i)

), for

i = 1. . . m. In this formulation, V

(i)

represents the in-

put sequence derived from video data, and z

(i)

denotes

the class label corresponding to the direction of the

shot. The classiﬁcation task involves three possible

categories: z

(i)

∈ {R, C, L} corresponding to Right,

Center, and Left from the goalkeeper’s perspective.

The objective is to learn the model parameters θ that

minimize the cross-entropy loss over the dataset:

J(θ) = −

∑

i=1

∑

k=1

p(z

(i)

= k) log(p(ˆz

(i)

= k)), (1)

where n = 3 is the number of output classes, p(z

(i)

k) is the ground truth probability for class k, and

p(ˆz

(i)

= k) is the predicted probability for class k for

instance i.

3.2 Gait Embedding

To extract motion features relevant to penalty kick-

ers, input sequences derived from video data V

(i)

are

ﬁrst converted into silhouette sequences, focusing ex-

clusively on the kicker (see Figure 1). Neither the

ball nor the goalkeeper is considered, and the only

visible silhouette in each sequence corresponds to the

player performing the kick. Detection and tracking

are performed using YOLOv8x-pose-p6 (Jocher et al.,

2023) and Bot-SORT (Aharon et al., 2022), while

high-precision silhouettes are obtained with SAMU-

RAI (Yang et al., 2024).

Each silhouette sequence is then processed using

a pre-trained gait recognition model B

GAIT

, trained

on large-scale public datasets such as CASIA-B (Yu

et al., 2006), OUMVLP (Takemura et al., 2018), and

GREW (Zhu et al., 2021). Rather than producing

a temporal embedding, B

GAIT

encodes the sequence

into a set of spatial gait descriptors corresponding

to vertically partitioned body regions. This structure

captures local motion characteristics across the body

from top to bottom. The embedding output is deﬁned

as:

= B

GAIT

(i)

) ∈ R

P×D

where P denotes the number of spatial body parti-

tions (e.g., P = 62 using Horizontal Pyramid Pool-

ing), and D is the embedding dimension per region.

To improve generalization and comparability across

samples, the embeddings are standardized using the

training set mean µ

train

and standard deviation σ

train

followed by L2 normalization:

[p] =

[p]− µ

train

[p] =

[p]

∥

[p]∥

The resulting normalized embedding matrix

∈

P×D

is then passed to a lightweight classiﬁcation

model.

Although

is not a temporal sequence, we treat

the vertical ordering of body parts (from head to foot)

as a structured sequence to capture spatial depen-

dencies. By applying a Recurrent Neural Network

(RNN) such as an LSTM over the P body partitions,

the model can learn hierarchical spatial interactions

across regions (e.g., how lower-body motion relates

to upper-body posture). This sequential processing

allows the classiﬁer to aggregate global pose infor-

mation while remaining sensitive to subtle localized

variations in movement style.

3.3 Shot Direction Classiﬁcation

A simple temporal classiﬁcation model is employed

to evaluate the predictive value of gait embeddings

for penalty kick direction. The objective is to learn

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

144

a mapping from the normalized gait sequence

∈

T ×D

to the shot label z

(i)

∈ {right, center, left}.

The model is deﬁned as a composition of standard

neural network layers:

(

) = Softmax



· ReLU



· LSTM

(

)

+ b



+ b



(2)

Where LSTM

(·) denotes a unidirectional LSTM

layer with 32 hidden units, W

∈ R

32×16

and W

∈

16×3

are weight matrices of the fully connected lay-

ers, and b

∈ R

and b

∈ R

are the correspond-

ing bias vectors. Softmax(·) converts the output logits

into a categorical distribution over the three classes.

This lightweight architecture is deliberately cho-

sen to avoid overﬁtting and isolate the gait embed-

dings’ representational quality. By limiting model ca-

pacity, performance differences across backbone net-

works can be more conﬁdently attributed to the dis-

criminative power of the extracted features, rather

than architectural complexity. Additionally, the

model’s transparent structure and low computational

cost make it suitable for fast benchmarking and itera-

tive experimentation.

4 EXPERIMENTAL SETUP

Gait Backbones. Silhouette-based gait recognition

methods aim to extract discriminative motion features

from binary human outlines to characterize walk-

ing behavior. For this study, we adapt a range of

representative architectures, including GaitBase (Fan

et al., 2023), GLN Phase 1 and 2 (Hou et al., 2020),

GaitGL (Lin et al., 2021), GaitPart (Fan et al., 2020),

and GaitSet, for the task of predicting penalty kick

direction. Though all these models process silhou-

ette sequences, their internal mechanisms differ sig-

niﬁcantly. GaitSet treats sequences as sets of inde-

pendent frames, using pooling operations to summa-

rize features over time, but it does not explicitly en-

code spatial continuity. GaitPart enhances this design

by focusing on horizontal body partitions through Fo-

cal Convolution, capturing localized motion patterns

but introducing potential sensitivity to pose misalign-

ment. GaitGL extends the modeling capacity by in-

tegrating global and part-based branches, alongside

3D convolutions, to extract joint spatial-temporal fea-

tures. While more expressive, its increased com-

plexity may limit performance consistency in real-

world applications. The GLN models adopt a la-

tent grouping mechanism. In a ﬁrst phase (Phase

1), grouped feature representations are built, and in

a second phase (Phase 2), progressive reﬁnement lay-

ers are introduced to increase representation granu-

larity across the network. Lastly, GaitBase provides

a deeper residual network baseline demonstrating the

effectiveness of capacity and depth without additional

architectural innovations.

As described in Section 3, the models were eval-

uated using embeddings trained on three benchmark

datasets: OU-MVLP, CASIA-B, and GREW. OU-

MVLP provides large-scale indoor sequences under

uniform conditions, while CASIA-B introduces con-

trolled variability through clothing, carrying objects,

and multi-view setups. Captured in the wild, GREW

presents more realistic challenges, such as occlusion

and lighting variation, which better reﬂect our target

domain. Not all backbones were trained on every

dataset, as architectural complexity and data variabil-

ity required careful pairing to ensure feasible training

and reliable feature extraction.

Dataset Collection and Filtering. The dataset

was constructed from publicly available footage span-

ning international matches, professional leagues, and

highlight compilations. Metadata about match level

was manually checked where available. Inspired by

the data acquisition strategies highlighted in prior

sports vision research (Thomas et al., 2017), the col-

lection focused on maximizing visual diversity re-

garding pose dynamics, camera distance, and illu-

mination. A targeted search using terms such as

“penalty-kick shootout” yielded a collection of raw

video clips, each manually trimmed to retain only the

relevant sequence, from the start of the run-up to the

outcome of the kick (see Figure 2).

To ensure temporal consistency and viewpoint

suitability, only clips recorded from optimal angles

(typically side or diagonal views of the kicker) and

with sufﬁcient temporal resolution (minimum of 64

frames) were retained. This ﬁltering process yield

to a dataset to 432 penalties. Each clip was anno-

tated with a shot direction label, where class 0 corre-

sponds to shots aimed left, 1 to center, and 2 to right.

The ﬁnal label distribution was imbalanced, with 209

samples in class 0, 66 in class 1, and 157 in class

2, which reﬂects real-world tendencies and must be

accounted for during model training and evaluation.

Goalkeepers appearing in the footage were not the fo-

cus of analysis but were used for the human baseline

by recording their initial dive direction.

Implementation details. The corresponding gait

embeddings were structured into ﬁxed-length se-

quences per penalty clip for each gait backbone under

evaluation. To preserve class distribution, a consis-

tent 80/20 train-test split was applied using stratiﬁed

sampling based on the shot direction label. The same

Gait-Based Prediction of Penalty Kick Direction in Soccer

145

Time

Silhouettes

RGB Samples

... ... ... ...

Figure 2: Silhouette sample frames from an RGB penalty sequence showing the kicker during the run-up and kick. The

silhouette view is cropped to include only the kicker, excluding the goalkeeper, to focus on the kicker’s motion patterns.

Table 1: Performance comparison of gait models and baselines for penalty kick direction classiﬁcation. Accuracy and

weighted recall are equal in this setting due to the use of single-label, multi-class classiﬁcation with one prediction per

sample. The goalkeeper baseline (human decision) is highlighted in gray, while the computational baseline (GaitBase) is

highlighted in blue.

Model Pretrained Dataset Accuracy Precision F1-Score

Goalkeeper N/A 46.0% N/A N/A

GaitBase (Fan et al., 2023) OUMVLP 48.3% 45.2% 46.4%

GaitBase (Fan et al., 2023) CASIA-B 51.7% 50.7% 51.1%

GaitPart (Fan et al., 2020) OUMVLP 51.7% 54.3% 51.1%

GaitPart (Fan et al., 2020) GREW 57.5% 57.3% 57.3%

GaitSet (Chao et al., 2018) OUMVLP 50.6% 50.8% 50.6%

GaitSet (Chao et al., 2018) CASIA-B 54.0% 52.4% 52.6%

GaitSet (Chao et al., 2018) GREW 52.9% 54.8% 52.4%

GLN Phase 1 (Hou et al., 2020) CASIA-B 56.3% 55.9% 55.5%

GLN Phase 2 (Hou et al., 2020) CASIA-B 52.9% 52.1% 52.0%

GaitGL (Lin et al., 2021) OUMVLP 54.0% 54.9% 53.9%

GaitGL (Lin et al., 2021) CASIA-B 56.3% 52.9% 53.7%

GaitGL (Lin et al., 2021) GREW 58.6% 49.9% 53.9%

split was reused across all models to ensure repro-

ducibility, with index mappings stored and reloaded

as needed. To account for variability due to model

initialization and training dynamics, each experiment

was repeated ﬁve times, and the resulting perfor-

mance metrics were averaged to provide a robust es-

timate of model effectiveness.

The classiﬁcation network was trained for up to

200 epochs with early stopping, using mini-batch

size 32 and the Adam optimizer with default learning

rate parameters. Class imbalance in the labels was

addressed through weighted loss computation, using

class weights derived from label frequency. As de-

scribed in Section 3, the model architecture consisted

of a single LSTM layer with 32 hidden units, fol-

lowed by a dropout layer (rate 0.5), a ReLU-activated

dense layer of size 16, and a softmax output layer with

three units corresponding to the directional classes.

Performance was measured using accuracy, precision,

and F1-Score. Accuracy and weighted recall give the

same result here because the model makes one pre-

diction for each video, and each video has only one

correct answer. Both metrics measure the same since

we count how many predictions are correct overall.

Baselines. Two types of baselines were consid-

ered to contextualize the performance of gait-based

models. The ﬁrst is a human decision baseline derived

from the goalkeeper’s initial dive direction. Impor-

tantly, this does not indicate whether the goalkeeper

successfully stops the shot. Instead, it simply reﬂects

the goalkeeper’s direction to dive, as recorded in the

dataset. In some cases, goalkeepers may initiate their

dive after the ball has already been struck. Therefore,

this baseline can be seen as a best-case scenario for

human anticipation, assuming access to all available

cues before the shot is taken.

In addition to the human benchmark, a compu-

tational baseline was established using a lightweight

gait recognition model commonly referred to as Gait-

Base. This architecture was selected due to its sim-

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

146

plicity, reproducibility, and solid performance across

controlled and unconstrained environments. Prior

work has shown that even minimalistic gait encoders

can yield competitive representations (Fan et al.,

2023), making GaitBase a strong reference point

for evaluating the added value of more sophisticated

backbone designs in the context of penalty kick anal-

ysis.

5 EXPERIMENTAL EVALUATION

Evaluating various gait recognition models for

penalty kick direction classiﬁcation reveals interest-

ing patterns across training datasets and model ar-

chitectures (see Table 1). Notably, the highest-

performing model overall was GaitGL trained on

GREW, achieving an accuracy of 58.6% and an F1

score of 53.9%. This suggests that models trained

on more diverse and in-the-wild datasets like GREW

may better generalize to the natural variability present

in broadcast soccer footage. GREW likely captures

broader pose, scale, and environmental variation com-

pared to more controlled datasets like OUMVLP or

CASIA-B, which may contribute to its superior trans-

fer performance.

Across the board, models trained on CASIA-

B tended to perform more consistently than those

trained on OUMVLP. CASIA-B’s multi-view setup

and moderate variability might offer a balance be-

tween structured feature learning and generalization

to new domains. For example, GLN Phase 1 (trained

on CASIA-B) achieved a strong F1 score of 55.5%, ri-

valing GaitGL’s performance on GREW. GaitGL and

GaitSet trained on CASIA-B also performed reliably,

with F1 scores of 53.7% and 52.6% respectively. This

trend highlights that while CASIA-B is a more con-

trolled dataset, its design may still support domain

transfer reasonably well for motion-based tasks.

In contrast, models trained on OUMVLP con-

sistently yielded lower performance, despite the

dataset’s large scale. GaitPart, GaitSet, and GaitGL

trained on OUMVLP all clustered around mid-range

F1 scores, achieving 51.1%, 50.6%, and 53.9% re-

spectively. The baseline model trained on OUMVLP

lagged behind at 46.4% F1, suggesting that size alone

does not guarantee effective transfer. OUMVLP’s

treadmill-based gait recordings may lack the natu-

ralistic movement patterns found in soccer approach

runs, reducing their representational relevance for this

task.

In terms of design, GaitGL emerged as one of

the most robust backbones across different training

domains. It achieved top-tier results on GREW and

CASIA-B and held up respectably on OUMVLP. Its

performance consistency indicates a strong capac-

ity for encoding discriminative motion dynamics in

downstream classiﬁcation tasks. GaitPart, although

effective in surveillance-style gait recognition, per-

formed moderately in this setup, likely due to its local

part-based focus, which might miss out on full-body

motion nuances relevant to shot prediction.

GLN also showed promise, especially in its

Phase 1 variant trained on CASIA-B. It matched

GaitGL in accuracy and demonstrated solid precision

and recall values, supporting its suitability for ﬁne-

grained action prediction. Phase 1 corresponds to

an intermediate checkpoint in the training of GLN,

where the backbone is already capable of extracting

meaningful gait representations but has not yet under-

gone full optimization for identity recognition. In-

terestingly, the Phase 2 variant, representing the ﬁnal

stage of training, underperformed slightly, with an F1

score of 52.0%. This performance drop may suggest

that the additional training in Phase 2 biases the model

more toward identity-speciﬁc features, potentially at

the expense of general motion cues relevant to ac-

tion understanding. Meanwhile, GaitSet exhibited a

balanced yet unremarkable proﬁle across all datasets,

suggesting that while effective, its aggregation-based

design may lack the temporal expressiveness needed

for predicting dynamic actions like kicks.

Lastly, the baseline models served as important

reference points. The CASIA-B-trained baseline out-

performed the OUMVLP version by a notable mar-

gin (F1: 51.1% vs. 46.4%), reinforcing that training

data characteristics critically shape downstream per-

formance. While none of the baselines matched the

top-performing models, their inclusion is crucial for

interpreting the added value of more complex archi-

tectures. The results demonstrate that advanced gait

models, especially GaitGL, GaitPart, and GLN, can

extract semantically meaningful motion features for

predictive tasks beyond identity recognition.

Error Analysis. Figure 3 presents the confu-

sion matrices for GaitGL and GaitPart trained on

the GREW dataset, revealing distinct misclassiﬁca-

tion patterns despite their similar overall accuracies

(58.6% for GaitGL and 57.5% for GaitPart). GaitGL,

while slightly more accurate, exhibits a pronounced

bias toward predicting the Right class. Notably, it

misclassiﬁes 69.2% of actual Center kicks and 40.6%

of Right kicks as Right, failing entirely to predict the

Center class. This indicates potential overﬁtting to di-

rectional features dominant in right-sided kicks, pos-

sibly stemming from imbalances in pose or silhouette

orientation during the run-up phase. Moreover, center

kicks exhibit less exaggerated lateral body motion and

Gait-Based Prediction of Penalty Kick Direction in Soccer

147

Figure 3: Normalized confusion matrices for models trained

on GREW. Each matrix shows prediction performance

across left, center, and right shot directions.

subtler preparatory cues than left or right shots. These

traits may lead GaitGL to underrepresent or overlook

the ﬁne-grained temporal features critical for distin-

guishing center shots, especially in penalty scenarios

where deceptive uniformity is typical.

In contrast, GaitPart demonstrates more balanced

predictions across all three classes. While its con-

fusion matrix still reﬂects a tendency to overpredict

Right for ambiguous cases, it is the only model among

the two that assigns non-zero predictions to the Cen-

ter class (38.5% for true center kicks). Additionally,

GaitPart maintains a better trade-off between preci-

sion and recall (57.3% and 57.5%), indicating that

it is less prone to skewed predictions and captures

a broader range of gait dynamics. This is particu-

larly evident in its handling of true Right kicks, where

it correctly classiﬁes 56.2%, compared to 59.4% by

GaitGL, but with fewer extreme misclassiﬁcations.

These observations suggest that while GaitGL

may achieve marginally higher accuracy by conﬁ-

dently predicting dominant classes, it lacks nuance

in handling more ambiguous gait patterns, especially

those associated with central shot directions. Gait-

Part, despite its slightly lower accuracy, appears to

better generalize across class boundaries. This trade-

off highlights an important consideration: models

with higher overall accuracy may still perform poorly

on minority or difﬁcult-to-classify cases, and con-

fusion matrix analysis is essential for understanding

these hidden weaknesses.

6 CONCLUSIONS

This study explored gait-based representations for

predicting penalty kick direction in soccer, extending

gait recognition beyond its usual role in identity ver-

iﬁcation. We used pre-trained gait backbones and a

lightweight temporal classiﬁer to assess whether mo-

tion patterns during the run-up could reliably indicate

shot direction. To support this, we curated a new

dataset of broadcast penalty sequences and evaluated

the transferability of gait embeddings across various

architectures and training domains.

Results showed that both the choice of backbone

and training data signiﬁcantly impact performance.

Models trained on GREW, an in-the-wild dataset,

performed best overall, with GaitGL achieving the

highest accuracy and F1 score. However, this came

with strong class bias, particularly toward right-side

predictions. GaitPart, though slightly less accurate,

yielded more balanced results across all directions, in-

cluding the underrepresented center. This highlights

the need to go beyond global metrics and consider

per-class behavior in imbalanced tasks.

These ﬁndings suggest that gait encodes meaning-

ful information about players’ motor intentions, even

in dynamic, high-pressure contexts like penalty kicks.

Our approach builds on recent work using Human

Action Recognition (HAR) models for penalty anal-

ysis (Freire-Obreg

on et al., 2025), which focus on

cropped segments like the run-up or kick. In contrast,

by preserving the entire sequence, our model captures

a broader range of kinematic and contextual cues.

While this may limit some traditional gait applica-

tions, such as short-cycle identiﬁcation, it enables ex-

ploration of richer temporal patterns. Notably, despite

departing from standard gait usage, the results show

that extended motion cues carry predictive value. This

opens the door to personalized modeling of penalty

direction, potentially tailored to each player’s unique

movement signature.

This research paves the way for practical sports

applications, such as decision-support tools, player

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

148

diagnostics, and automated video analysis. Incor-

porating features like foot orientation, ball trajec-

tory, or goalkeeper behavior could boost accuracy and

contextual understanding. Future work may explore

multi-modal fusion or adapt gait models to better cap-

ture sport-speciﬁc movement patterns.

ACKNOWLEDGEMENTS

This work is partially funded funded by

project PID2021-122402OB-C22/MICIU/AEI

/10.13039/501100011033 FEDER, UE and by the

ACIISI-Gobierno de Canarias and European FEDER

funds under project ULPGC Facilities Net and Grant

EIS 2021 04.

REFERENCES

Aharon, N., Orfaig, R., and Bobrovsky, B.-Z. (2022). BoT-

SORT: Robust associations multi-pedestrian tracking.

arXiv preprint arXiv:2206.14651.

Artiles, J., Hern

andez-Sosa, D., Santana, O., Lorenzo-

Navarro, J., and Freire-Obreg

on, D. (2024). Clas-

sifying soccer ball-on-goal position through kicker

shooting action. In Proceedings of the 13th Inter-

national Conference on Pattern Recognition Appli-

cations and Methods - ICPRAM, pages 79–90. IN-

STICC, SciTePress.

Chao, H., He, Y., Zhang, J., and Feng, J. (2018). Gaitset:

Regarding gait as a set for cross-view gait recognition.

In AAAI Conference on Artiﬁcial Intelligence.

ChyronHego (2017). 5 reasons you need tracab player

tracking. Accessed 12 Feb 2017.

Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., and Yu, S.

(2023). Opengait: Revisiting gait recognition towards

better practicality. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion (CVPR).

Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang,

Y., Li, Q., and He, Z. (2020). Gaitpart: Temporal part-

based model for gait recognition. 2020 IEEE/CVF

Conference on Computer Vision and Pattern Recog-

nition (CVPR), pages 14213–14221.

Figueroa, P., Leite, N., and Barros, R. (2006). Tracking soc-

cer players aiming their kinematical motion analysis.

Computer Vision and Image Understanding, 101:122–

135.

Freire-Obreg

on, D., Lorenzo-Navarro, J., Santana, O. J.,

Hern

andez-Sosa, D., and Castrill

on-Santana, M.

(2022). Towards cumulative race time regression in

sports: I3d convnet transfer learning in ultra-distance

running events. In 2022 26th International Confer-

ence on Pattern Recognition (ICPR), pages 805–811.

Freire-Obreg

on, D., Santana, O. J., Lorenzo-Navarro,

J., Hern

andez-Sosa, D., and Castrill

on-Santana, M.

(2025). Predicting Soccer Penalty Kick Direction Us-

ing Human Action Recognition. In 2025 23rd Interna-

tional Conference on Image Analysis and Processing

(ICIAP).

Hou, S., Cao, C., Liu, X., and Huang, Y. (2020). Gait Lat-

eral Network: Learning Discriminative and Compact

Representations for Gait Recognition. In European

Conference on Computer Vision, page 382–398.

Innovations, H.-E. (2017). Hawk-eye goal line technology.

Accessed 12 Feb 2017.

Jocher, G., Chaurasia, A., and Qiu, J. (2023). Ultralytics

yolov8.

Kapela, R., McGuinness, K., Swietlicka, A., and O’Connor,

N. (2014). Real-time event detection in ﬁeld sport

videos. In Moeslund, T., Thomas, G., and Hilton,

A., editors, Computer Vision in Sports, chapter 14.

Springer.

Lin, B., Zhang, S., and Yu, X. (2021). Gait recognition via

effective global-local feature representation and local

temporal aggregation. IEEE/CVF International Con-

ference on Computer Vision (ICCV), pages 14628–

14636.

Liu, J. and Carr, P. (2014). Detecting and tracking sports

players with random forests and context-conditioned

motion models. In Moeslund, T., Thomas, G., and

Hilton, A., editors, Computer Vision in Sports, chap-

ter 6. Springer.

Maksai, A., Wang, X., and Fua, P. (2016). What players

do with the ball: A physically constrained interaction

modeling. In Proceedings of the IEEE Conference on

Computer Vision and Pattern Recognition (CVPR).

Manaﬁfard, M., Ebadi, H., and Moghaddam, H. (2016). A

survey on player tracking in soccer videos. Computer

Vision and Image Understanding. Submitted for Spe-

cial Issue on Computer Vision in Sports.

Takemura, N., Makihara, Y., Muramatsu, D., Echigo, T.,

and Yagi, Y. (2018). Multi-view large population gait

dataset and its performance evaluation for cross-view

gait recognition. IPSJ Transactions on Computer Vi-

sion and Applications, 10:1–14.

Thomas, G., Gade, R., Moeslund, T. B., Carr, P., and Hilton,

A. (2017). Computer vision for sports: Current ap-

plications and research topics. Computer Vision and

Image Understanding, 159:3–18.

Yang, C.-Y., Huang, H.-W., Chai, W., Jiang, Z., and Hwang,

J.-N. (2024). SAMURAI: Adapting segment any-

thing model for zero-shot visual tracking with motion-

aware memory.

Yu, S., Tan, D., and Tan, T. (2006). A framework for

evaluating the effect of view angle, clothing and car-

rying condition on gait recognition. 18th Interna-

tional Conference on Pattern Recognition (ICPR’06),

4:441–444.

Zhu, Z., Guo, X., Yang, T., Huang, J., Deng, J., Huang, G.,

Du, D., Lu, J., and Zhou, J. (2021). Gait recognition

in the wild: A benchmark. 2021 IEEE/CVF Interna-

tional Conference on Computer Vision (ICCV), pages

14769–14779.

Gait-Based Prediction of Penalty Kick Direction in Soccer

149