Vision-Based System for Tennis Stroke and Error Classification

Tennis practice without live guidance often cements small technical flaws an early wrist bend on the backhand, a lop-sided slice, or a hurried serve that over time become bad habits and even lead to injuries like tennis elbow. While one-on-one lessons can correct these, they come at a steep price. This work presents a user-friendly, camera-only feedback system powered by modern computer vision and deep learning. We began with a simple template matcher on pose landmarks, reaching 71.6 % accuracy, then built three stronger models: a CNN + LSTM network over landmark sequences (81 % accuracy), a YOLO-based CNN detector on full video frames (94.5 %), and a Vision Transformer in Keras (92.1 %). An ensemble of all three boosts overall accuracy to 95 %. Training drew on the public THETIS dataset and our own set of 300 annotated "faulty" clips—captured in varied real-world settings from indoor courts to sunlit outdoor walls, with players spanning complete beginners to club competitors. This lightweight system runs in real time on any webcam feed, enabling affordable, on-demand stroke feedback at home, in the club, or anywhere in between.