Abstract
This system utilizes an STM32F103 as the main control chip to measure instantaneous acceleration and shot angles via an MPU-6050 sensor. Motion data was collected from professional players performing instep shots to train a custom Convolutional Neural Network (CNN) model. This model is deployed to a mobile application, which receives real-time data from the wearable device via an HC-05 Bluetooth module. The system evaluates shooting technique, provides visualization of training data, and offers pertinent suggestions for improvement. Designed for ease of use, the system serves as an effective tool for beginners to practice and for coaches to monitor and guide player development.
Introduction
While soccer is a global passion, professional coaching remains a luxury. Beginners often lack the scientific guidance needed to improve their shooting mechanics, and private coaches are financially out of reach for many families. This research identifies a gap in the "smart wear" market: while deep learning has revolutionized image and speech recognition, its application in wearable sports tech has been limited by a lack of high-quality motion datasets. This paper aims to bridge that gap by creating a system that captures the "gold standard" of professional movement and uses it to evaluate and guide amateur players.
System Architecture
The system is built on an Internet of Things (IoT) architecture consisting of three main components: hardware sensing, a CNN-based processing model, and a mobile interface.
Hardware Sensing
An STM32F103 (ARM Cortex-M3) serves as the central processing unit, interfacing with an MPU-6050 six-axis sensor. This unit is worn by the player on the leg to measure instantaneous acceleration and angular velocity during a shot. Data is transmitted from the wearable device to a smartphone using an HC-05 Bluetooth module.
The CNN Model and Dataset
The core intelligence is a Convolutional Neural Network (CNN) designed to extract features from motion data. The architecture includes two convolutional layers for feature extraction, two pooling layers to reduce dimensionality, a fully connected layer, and a final classification layer. To make the 1D sensor data compatible with the CNN, raw time-series data is converted into a matrix format before feeding it into the model.
The authors developed a custom dataset by collecting motion data from a large number of professional players performing the instep shot. This professional data serves as the benchmark for the "excellent" classification in the model.
Results and User Interface
The system categorizes shots into levels such as Poor, Good, and Excellent based on the CNN's output. Training results established specific ranges for high-quality shots:
- Optimal Launch Angle: The model identifies an angle between 48.4° and 57.5° as "excellent." Angles below 26.7° or above 57.6° are flagged as poor.
- Acceleration Benchmarks: "Excellent" shots showed acceleration values between 113.9 and 162.4 m/s². Values above 162.5 m/s² are categorized as "good," suggesting a potential loss of control at extreme power levels.
The mobile app visualizes the shot data and provides feedback. Beyond mechanical analysis, the app suggests a "total athlete" approach by recommending diet (e.g., specific foods like apples) and supplementary exercises to help beginners practice effectively.
Conclusion and Future Work
The current system successfully digitizes professional technique to provide an objective benchmark for amateurs. However, the authors acknowledge that the training effect is currently limited by dataset size. Future work will focus on expanding the dataset with more "classic shot data" to improve robustness, increasing classification accuracy for professional-grade use, and broadening the scope to include other techniques like headers and volleys.
Summary updated