TTSWING Player Identification Using Swing Embeddings
Biometric identification of table tennis players using inertial sensor data and learned swing embeddings.
Gallery
Notes
Overview
This project investigated whether table tennis players can be uniquely identified using swing biomechanics captured by inertial sensors embedded in racket grips.
Description
Using the TTSWING dataset, each stroke was represented by 36 engineered features derived from 9 axis IMU data across time and frequency domains. A strong non parametric baseline using k nearest neighbours achieved approximately 80 percent accuracy across 93 players, demonstrating that swing dynamics contain distinctive biometric signatures.
However, further analysis revealed limitations in scalability and fairness, as k-NN performance was biased toward players with more recorded strokes and required storing the full training set.
To address this, the project explored metric learning using a Siamese network trained with triplet loss to produce compact 64 dimensional embeddings. This backbone was fine tuned with a supervised classification head using balanced sampling, weighted loss, dropout, batch normalisation, and a one cycle learning rate schedule. The final model achieved approximately 68 percent test accuracy, with significantly improved robustness for under represented players.
While k-NN remained superior in raw accuracy, the learned embeddings offered clear advantages in generalisation, interpretability, and deployment potential. The project received top band marks.