Tennis shot recognition (B.Tech thesis)

PyTorch TensorFlow MoveNet CNN RNN LSTM Attention
GitHub External link

Problem

Classify tennis shots from video using pose trajectories instead of brittle pixel-only heuristics.

Approach

  • MoveNet for human pose estimation on frames.
  • Sequential models (RNN, GRU, LSTM, BiLSTM) with attention and CNN encoders to capture motion patterns.
  • Built a custom dataset and an end-to-end ML pipeline aimed at near–real-time coaching and analytics.

Outcome

~94% accuracy on the thesis task — with a path toward real-world deployment constraints (latency, data drift, and labeling).