Literature Review on Machine Learning

in Football Player Detection and

Analysis
By Shubham Tanna E026

Introduction
Machine learning (ML) is rapidly transforming sports analytics, particularly in football, where
advanced algorithms allow for automated player detection, tracking, and performance evaluation.
This literature review examines three research papers focusing on diverse aspects of ML in football:
a deep convolutional neural network (CNN)-based detection model, an upgraded YOLOv5
architecture, and an OpenCV-based optical tracking approach. These studies provide insight into the
capabilities and challenges of using ML for real-time sports analytics, an area with significant
potential for improving game strategies, performance insights, and enhancing viewer experience.
The review covers the approaches, key findings, challenges, and future directions suggested by
each study.

Literature Review

Paper 1: Deep Learning-Based Football Player Detection in

Videos by Tianyi Wang and Tongyan Li (2022)

Problem Statement
This research addresses a fundamental challenge in real-time sports analytics: accurately detecting
and tracking football players within video footage, particularly under varied and often unpredictable
conditions. Traditional methods have struggled to balance high accuracy with computational
efficiency, especially when dealing with complex backgrounds, diverse video resolutions, and the
need for real-time processing on portable devices. The study’s primary objective is to develop a
model capable of reliably detecting football players with minimal latency, irrespective of the
resolution or dynamic nature of the background, making it accessible for use in applications such as
sports analytics, real-time broadcasting, and tactical performance analysis.

Techniques Used
To tackle the challenges identified, the authors implemented a deep convolutional neural network
(CNN) model enhanced with several advanced techniques for effective feature extraction and
efficient computation:

● Feature Pyramid Networks (FPN): The model integrates FPNs, which help in extracting
multi-scale features from input images. This is especially useful in sports analytics where
player sizes and distances from the camera vary significantly within a single video, making it
essential to capture fine details for players further away while maintaining clarity for closer
ones.
● Residual Connections: To improve gradient flow and enable deeper network architectures
without degradation, the model incorporates residual connections. These connections help
the network learn complex patterns and reduce issues like vanishing gradients, which often
hinder deep learning in video analysis tasks.
● Convolutional Blocks and Activation Functions: The architecture consists of five
convolutional blocks, each fine-tuned with batch normalization and leaky ReLU activation.
Batch normalization helps stabilize the learning process, and leaky ReLU addresses the
dying ReLU problem, maintaining sensitivity to a wide range of spatial features. These
components are crucial for ensuring that the network can operate effectively across varying
spatial resolutions, a common challenge in video-based applications.
● Efficient Model Design for Portability: The network is designed to be lightweight,
optimizing both memory and computational requirements. This lightweight nature makes the
model suitable for deployment on portable devices, where high computational power may not
be available, thus expanding its usability to on-field and mobile sports analytics applications.

Key Findings
The model was evaluated on the ISSIA-CNR dataset, a widely recognized benchmark dataset for
sports analytics, particularly football. Key metrics and findings from this evaluation include:

• High Accuracy and Precision: The model achieved a mean average precision (mAP) of
0.915, indicating high accuracy in detecting players across varied scenarios. This level of
precision positions the model as a competitive alternative to traditional detection frameworks,
particularly YOLO, which is known for its efficiency but often sacrifices precision in crowded or
complex scenes.

• Computational Efficiency: Compared to conventional models like YOLO, this CNN-FPN

model demonstrated superior computational efficiency, processing frames faster and using less
computational power. This efficiency makes it ideal for applications where real-time analysis is
required, such as live sports broadcasting, without compromising detection accuracy.
 • Robustness Across Resolutions: The model maintained its performance across different
video resolutions, a critical capability for sports analytics where high-definition and
standard-definition feeds are often mixed. This adaptability showcases the model’s versatility in
diverse media environments, which can be a significant asset for broadcasters and analysts
working with varying data quality.

Limitations
Despite its strengths, the model has certain limitations, which the authors address as areas for future
improvement:

• Detection of Small or Occluded Players: The model encounters challenges when

detecting small players (those farther from the camera) or partially occluded players in crowded
areas. The FPN helps with multi-scale detection but does not fully resolve issues related to
occlusion, which often results in missed detections in crowded or close-contact gameplay.

• Complex Backgrounds and Dynamic Lighting: In scenes with highly complex

backgrounds or fluctuating lighting conditions (e.g., changing shadows or artificial lighting shifts),
the model’s performance drops, impacting the consistency of detection. This issue is common in
video-based applications where backgrounds and lighting are constantly changing, suggesting a
need for further advancements in background modeling and dynamic lighting adaptation.

• Reliability in Real-World Scenarios: While the model performs well in controlled or

semi-controlled environments, it may require additional training or fine-tuning to handle the
unpredictability of real-world footage, such as crowd interference or varying field conditions.

Paper 2: Football Player Tracking and Performance Analysis

Using OpenCV Library by Baydaa M. Merzah et al. (2024)

Problem Statement
This study addresses the high costs and limitations associated with traditional player tracking
systems such as GNSS (Global Navigation Satellite Systems) and LPS (Local Positioning Systems),
which are widely used in professional sports to monitor player metrics like speed, position, and
movement patterns. The authors propose an alternative approach that leverages OpenCV’s
computer vision capabilities for optical tracking. By using camera footage alone, this model aims to
provide an affordable, accessible solution to track football players in real time, making it particularly
useful for training scenarios where high-cost equipment may be impractical or unavailable.

Techniques Used
The study employs the following key methods to accomplish real-time player tracking and analysis:
 • CSRT (Channel and Spatial Reliability Tracking) Algorithm: The CSRT algorithm in
OpenCV, which is known for its robustness and accuracy, was chosen as the primary method for
tracking player movement frame by frame. CSRT allows for more stable tracking of objects by
updating the bounding box, which makes it resilient to moderate occlusions and some player
movements. This algorithm is particularly effective in maintaining track even if the object (player)
undergoes slight changes in appearance.

• Euclidean Distance for Positional Metrics: To compute essential metrics like player
position and movement speed, the authors use Euclidean distance calculations between
consecutive frames. This metric enables the model to track how far a player has moved within a
certain time interval, allowing for an estimation of speed and distance covered.

• Frame Segmentation for Player Isolation: To improve tracking accuracy and ensure that
the algorithm tracks individual players, frame segmentation is applied. This step is essential for
isolating specific players, reducing interference from other objects or players in the background.
By isolating players in each frame, the model minimizes errors and improves the reliability of
metrics like speed and position.

Key Findings
The OpenCV-based model demonstrated promising results, especially in terms of cost-effectiveness
and ease of use:

• High Accuracy in Short Segments: The model was found to perform well in short video
segments, reliably capturing player movements and position without the need for additional
hardware like sensors or GPS. This finding suggests that optical tracking through OpenCV can
serve as a viable alternative for training and analysis in environments with limited resources.

• Potential for Real-Time Applications: The OpenCV-based approach proved effective for
real-time tracking, which is beneficial for training scenarios where coaches and analysts need
immediate feedback on player movements and performance. The reliance solely on camera
footage provides flexibility, as it can be applied in various setups without specialized hardware.

• Applications in Training Scenarios: Due to its low cost and ease of use, this model offers
significant potential in non-professional or training environments, making advanced player
tracking accessible to a wider audience and enabling performance analysis without the logistical
challenges associated with GNSS or LPS systems.

Limitations
While the OpenCV approach offers many benefits, the study also identifies several limitations that
affect the model’s performance and scalability:

• Manual Region Selection: A key limitation is the need for manual selection of regions to
track individual players. This manual setup limits scalability and makes the model less suitable
for real-time applications in dynamic or larger environments where players frequently move
 across the frame. Without automation, this process is labor-intensive and may not be practical in
high-paced scenarios.

• Dependence on Video Quality: The accuracy of the tracking system is heavily dependent
on the quality of the video feed. Low-resolution or noisy video footage reduces the model’s
ability to accurately detect and track players, particularly when they are distant or partially
occluded. In high-density environments, such as crowded playing fields, video quality issues can
further reduce accuracy, impacting the overall reliability of performance metrics.

• Crowded Scene Limitations: In scenes with high player density or significant occlusion, the
model struggles to maintain accurate tracking. CSRT is generally robust but has limitations in
handling highly cluttered environments, which can lead to misinterpretation of movement
patterns or incorrect tracking results.

Paper 3: Enhancing the Performance and Accuracy in Real-Time

Football and Player Detection Using Upgraded YOLOv5
Architecture by Keyan Zhao (2024)

Problem Statement
This study focuses on optimizing the YOLOv5 model to address the specific challenges of real-time
football player detection, particularly under conditions involving occlusion and dynamic lighting.
Traditional YOLO models, though effective for general object detection, struggle to maintain high
accuracy and speed when dealing with fast-moving subjects in sports settings. The objective of the
paper is to enhance the YOLOv5 architecture to deliver faster, more precise player detection, with
practical applications in live broadcasting, real-time analytics, and automated sports analysis.

Techniques Used
The authors implement several key architectural upgrades to improve the baseline YOLOv5 model.
These enhancements target multi-scale feature extraction, model efficiency, and bounding box
precision to optimize the model for real-time sports analytics:

• SimSPPF (Simplified Spatial Pyramid Pooling – Fast): To capture features at multiple

scales, the model incorporates SimSPPF. This module performs spatial pyramid pooling,
enabling the model to recognize players of varying sizes and at different distances from the
camera. Multi-scale feature extraction is crucial in sports applications to detect players
accurately across various parts of the field, especially when players are at different distances
and angles from the camera.

• GhostNet Backbone: GhostNet is utilized in the backbone of the network to reduce the
model’s complexity without sacrificing its feature extraction capacity. GhostNet introduces “ghost
modules,” which allow for a more compact representation of features, ultimately lowering the
 computational load. This reduction in complexity enhances processing speed, making the model
more efficient and suitable for real-time applications where latency is a concern.

• Slim-Scale Detection Layer: A new slim-scale detection layer is added to refine bounding
box predictions, ensuring that player localization is precise, even in scenes with fast movement
and frequent changes in direction. This layer improves the model’s capability to adapt to player
movements and is particularly useful in sports scenarios where players frequently shift positions
within the frame.

Key Findings
The upgraded YOLOv5 model demonstrated substantial improvements in both accuracy and speed,
positioning it as a viable solution for real-time sports analytics:

• Enhanced Performance: The modified YOLOv5 achieved a 15% increase in mean Average
Precision (mAP) at an Intersection over Union (IoU) threshold of 0.5. This metric indicates a
significant boost in detection accuracy, essential for minimizing false positives and false
negatives in a dynamic environment like a football field.

• Improved Processing Speed: With the architectural enhancements, the model processes
video frames more quickly, reducing latency and making it suitable for live broadcasting. The
improvements in speed ensure that detection can occur in near real-time, allowing broadcasters
and analysts to access insights immediately as the game progresses.

• Versatile Application in Sports Broadcasting: The model’s adaptability to varying lighting

conditions and the capability to handle minor occlusions make it ideal for sports broadcasting,
where real-time analysis and player identification are critical. These improvements demonstrate
the model’s applicability not only in recording scenarios but also in live settings, where timely
data is essential.

Limitations
While the upgraded YOLOv5 model addresses several challenges, certain limitations remain,
particularly related to occlusion and crowded scenes:

• Handling Severe Occlusions: Despite the use of multi-scale feature extraction and the
GhostNet backbone, the model still encounters difficulties when players are severely occluded
by others. This limitation affects the model’s ability to accurately track individual players in
crowded or overlapping conditions, a common issue in team sports with high player density in
certain field areas.

• Distinguishing Overlapping Players: In situations where players are in close proximity or

partially overlap, the model struggles with accurately distinguishing them. This issue can lead to
errors in player identification and tracking, impacting the quality of data for analytics and
reporting.
 • Future Work Suggestions: The paper suggests exploring additional occlusion-handling
techniques in future studies. Potential solutions could include integrating temporal tracking
mechanisms or utilizing depth data, which could enhance the model’s capacity to differentiate
players in crowded scenarios.

Comparative Analysis
Model Complexity:
The upgraded YOLOv5 model (Paper 3) is the most complex, leveraging multi-scale feature
extraction and optimized layers for enhanced real-time performance. The CNN-based model (Paper
1) is simpler and lightweight, suitable for portable devices but not as robust in challenging scenes.
The OpenCV approach (Paper 2) is the simplest and most cost-effective but lacks deep learning’s
automatic adaptability.

Performance and Application Scope:

The YOLOv5 model outperforms in terms of speed and detection accuracy, making it ideal for live
broadcasts and large-scale applications. The CNN model is efficient but struggles in high-occlusion
scenarios, while the OpenCV-based model provides cost-effective tracking for smaller-scale or
controlled environments.

Real-Time Applicability:
The YOLOv5 model excels in real-time settings due to its computational efficiency and optimized
bounding box predictions. In contrast, the CNN and OpenCV models are limited in this regard—CNN
by processing speed and OpenCV by the need for manual setup.

Occlusion Handling
Only the YOLOv5 model includes modifications to handle occlusions, though it remains imperfect in
crowded conditions. The CNN model’s reliance on lower-resolution images limits its ability to
manage occlusion effectively, while the OpenCV model is susceptible to inaccuracies in crowded or
cluttered environments .
Trends and Challenges
Trends:
● Real-Time Processing: There is a growing demand for real-time ML models in sports,
evident from the focus on optimized architectures and feature extraction in the YOLOv5
model.
● Lightweight Architectures: Portable, lightweight models like the CNN in Paper 1 show an
increasing focus on efficient, device-compatible ML applications in sports.
● Cost-Effective Solutions: The use of OpenCV demonstrates the trend towards affordable,
scalable solutions for player tracking, especially in environments without high-tech
infrastructure.

Challenges:
● Occlusions and Overlapping Objects: A common issue across the papers is the difficulty
in detecting players during occlusions or overlapping. YOLOv5 enhancements address this,
yet it remains a limitation.
● Scalability in Dynamic Environments: Both the CNN and OpenCV models face scalability
issues—CNN in processing power, and OpenCV due to the requirement for manual
selection.
● Reliance on High-Quality Input: The OpenCV-based model’s accuracy is highly dependent
on video quality, indicating that advances in low-resolution detection are still needed.
● Real-Time Constraints: While the YOLOv5 model progresses in real-time capabilities, the
CNN and OpenCV models lag behind due to processing demands and manual setup needs.

Future Directions
● Enhanced Occlusion Handling: Future work could focus on hybrid models combining
optical flow with deep learning to improve occlusion handling, especially in crowded scenes.
● Resource-Efficient Architectures: For real-time performance, future research may explore
lighter, more efficient architectures that still maintain detection accuracy, suitable for
deployment on portable devices.
● Improving Low-Resolution Detection: Developing models capable of accurate player
detection from low-resolution or lower-quality video sources could broaden the application
scope of ML in sports analytics.
● Automated Segmentation in Tracking Models: Integrating automated segmentation
techniques in models like OpenCV-based tracking can improve scalability in dynamic,
real-time environments.
Conclusion
This literature survey highlights the growing role of ML in football player detection and tracking,
showcasing advancements and limitations in three key approaches. The YOLOv5-based model
excels in real-time performance and complex detection tasks, the CNN model offers a lightweight
alternative for simpler applications, and OpenCV tracking provides a cost-effective solution for
smaller-scale use cases. Each model shows promise for improving sports analytics but faces
challenges in handling occlusions, maintaining efficiency, and scaling to dynamic environments.
Future research in these areas will likely focus on creating more versatile, resource-efficient models
to meet the needs of diverse applications in sports analysis and live broadcasting.

References
1. Wang, Tianyi & Li, Tongyan. (2022). Deep Learning-Based Football Player Detection in Videos.
Computational Intelligence and Neuroscience. 2022. 1-8. 10.1155/2022/3540642.
2. Merzah, B.M., Croock, M.S., Rashid, A.N. (2024). Football player tracking and performance analysis
using the OpenCV library. Mathematical Modelling of Engineering Problems, Vol. 11, No. 1, pp.
123-132. https://doi.org/10.18280/mmep.110113
3. Zhao, Keyan. (2024). Enhancing the Performance and Accuracy in Real-Time Football and Player
Detection Using Upgraded YOLOv5 Architecture. International Journal of Computational Intelligence
Systems. 17. 10.1007/s44196-024-00565-x.