Sports activity (SA) recognition based on error correcting output codes (ECOC) and convolutional neural network (CNN)

Area of Science:

• * Sports science and biomechanics
• * Machine learning and artificial intelligence
• * Wearable sensor technology

Background:

• * Motion sensors are increasingly used for activity monitoring, particularly in sports.
• * Current sensor-based systems struggle to differentiate specific sports activities from daily routines.
• * Accurate sports activity recognition is crucial for athlete monitoring and training analysis.

Purpose of the Study:

• * To develop a novel machine learning model for enhanced sports activity recognition.
• * To accurately distinguish between various sports activities and everyday human movements.
• * To improve the precision and recall of sensor-based activity detection systems.

Main Methods:

• * Utilized an accelerometer and gyroscope attached to the foot for data collection.
• * Applied Short-Time Fourier Transform (STFT) for signal feature extraction.
• * Employed a Convolutional Neural Network (CNN) for motion characteristic analysis.
• * Used an Error Correction Output Code (ECOC) model for final activity classification.

Main Results:

• * The proposed model achieved a high accuracy of 99.71% in sports activity recognition.
• * Achieved precision of 99.72% and recall of 99.71% on the DSADS database.
• * Demonstrated superior performance compared to existing methods in distinguishing activities.

Conclusions:

• * The developed machine learning approach significantly improves sports activity recognition accuracy.
• * Foot-mounted sensors combined with STFT, CNN, and ECOC offer a robust solution.
• * This method provides a reliable tool for objective assessment in sports science and training.