Wearable sensing module for Table Tennis stroke detection

Sensor-based analytics is a science of capturing the essentials of human motion and has been extensively used in sports by attaching external sensors to players for deriving meaningful insights, thereby improving game-playing quality. This paper presents the work towards developing a wearable Body Sensor Network(BSN) for Table Tennis(TT) stroke detection through self-learned features. By capturing acceleration and rotation parameters using Inertial Measurement Unit (IMU) sensors mounted on the player forearm and shoulder while playing a TT stroke, a neural network has been trained to predict forehand and backhand shots. The model runs on Arduino Nano 33 BLE and achieves a real-time prediction accuracy of 90.7%. Also, a memory-efficient Tensorflow-based Quantization Aware Training(QAT) model, with up to 4 times reduction in model size, has been developed to demonstrate comparable prediction accuracy with size reduction.