CBAM-Enhanced CNN-LSTM with Improved DBSCAN for High-Precision Radar-Based Gesture Recognition.

Area of Science:

• Computer Science
• Signal Processing
• Artificial Intelligence

Background:

• Radar-based gesture recognition is vital for industrial and daily applications.
• Complex scenarios present challenges like clutter, similar gestures, and ambiguous features, limiting current algorithm performance.

Purpose of the Study:

• To propose a novel framework, CECL, for high-accuracy and robust radar-based gesture recognition.
• To enhance spatial-temporal feature extraction and clutter suppression for improved performance.

Main Methods:

• A Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture integrated with the Convolutional Block Attention Module (CBAM).
• Signal processing techniques including Blackman window for spectral leakage suppression, wavelet thresholding, and dynamic energy nulling for clutter reduction.
• An improved Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm for noise elimination.

Main Results:

• The CECL framework achieved an average accuracy of 98.33% in gesture classification, surpassing baseline models.
• Demonstrated excellent recognition performance across varying distances and angles, indicating enhanced robustness.

Conclusions:

• The proposed CECL framework effectively addresses challenges in radar-based gesture recognition.
• Achieves state-of-the-art accuracy and robustness, making it suitable for complex real-world applications.