LSTM Networks for Fine Gesture Classification

Resumo

Resumo

This study explores the application of long shortterm memory (LSTM) networks for gesture recognition via accelerometer and gyroscope data from smartwatches. We evaluate the impact of key parameters-window size ($2-4 \mathrm{~s}$), overlap $(5-15 \%)$, input signals (accelerometer-only vs. accelerometer-gyroscope), sampling frequency (25 Hz vs. 50 Hz, and low-pass filtering on the classification performance for 6 and 7 gesture classes. Our results demonstrate that optimal performance (test accuracy: 0.993, F1 score: 0.993) is achieved with a 4 -second window, $5 \%$ overlap, 6 input signals, 50 Hz frequency, and filtering. For 7 gestures, the best accuracy decreases to 0.888, highlighting increased complexity. Smaller windows (2 s) or accelerometer-only inputs significantly degrade performance, with accuracies below 0.85. The LSTM model outperforms traditional methods, such as random forest, particularly for larger gesture sets, while reducing the number of preprocessing steps and computational costs. These findings underscore the importance of parameter tuning and sensor fusion in achieving robust and energyefficient gesture recognition in wearable devices.

palavras-chave

Palavras-chave

gesture recognitionlong short-term memory networkssmartwatchesaccelerometergyroscopesensor fusionwindow sizesampling frequencylow-pass filteringwearable devicesparameter tuningenergy efficiency

links

Links e arquivos