紀要論文

Fine-Tuning Frame Range in mmWave Radar Point Clouds Generation for Sign Language Gesture Recognition with Bi-LSTM Attention

eng

24321192

https://doi.org/10.15002/00030603

法政大学大学院情報科学研究科

法政大学大学院紀要. 情報科学研究科編

19

1

6

2024-03-24

Version of Record

This research focuses on fine-tuning the frame range for point clouds generated by a millimeter wave radar in the context of American Sign Language (ASL) gesture recognition. The study employs an IWR6843AOP millimeter wave radar, a non-contact device, to capture fifteen ASL gestures, obtaining up to 800 samples per sign gesture. Initially, the frame range varies from 11 to 120. Training an LSTM model with this initial frame range yields an overall accuracy of 88.3% for 15 gestures. By fine-tuning the frame range to [30-54] using a histogram distribution, the accuracy of sign gesture recognition improves to 89.2%. Further enhancement is achieved by adopting a Bi-LSTM learning model, resulting in an accuracy of 91.7%, utilizing approximately 700 samples per gesture as the optimal sample size. The research employs two algorithms for fine-tuning the frame range and explores three learning models: LSTM, Bi-LSTM, and Bi-LSTM with an Attention mechanism. The obtained accuracies for sign gesture recognition are 89.2%, 91.7%, and 92.30%, respectively. This study demonstrates the efficacy of optimizing the frame range and leveraging Bi-LSTM models with attention mechanisms for improved ASL gesture recognition accuracy.

Article

AA12746425