Extraction of tackling motion of rugby players using machine learning and inertial sensors

Abstract

In the top league of rugby, players are worn with IMU sensors that record their kinematic characteristics, and they are searching for effective ways to utilize the data. We focus on tackling motions and aim to develop an AI that extracts and evaluates signals corresponding to tackling motions from the signals measured by inertial sensors. This study analyzed wearable IMU data and video data collected from official matches played by the NTT Docomo Red Hurricanes of the Top League in 2020-2021. The AI development software uses NNC (Neural Network Console) to create a CNN model. And AI is evaluated by obtaining learning curves and confusion matrices. In a study conducted in 2020-2021, training was performed using the developed AI with 73 pieces of teacher data and 32 pieces of validation data as input, and the correct response rate was 6.25%. This could be attributed to the small amount of teacher data and the recognition rate of each tackle waveform. In this report, we summarize the series of steps from the creation of teacher data to the development, execution, and evaluation of the AI.