MCT-CNN-LSTM: A Driver Behavior Wireless Perception Method Based on an Improved Multi-Scale Domain-Adversarial Neural...

Area of Science:

• Radar Signal Processing
• Machine Learning
• Automotive Safety

Background:

• Frequency-Modulated Continuous-Wave (FMCW) radar systems are widely used for driving behavior recognition.
• Deep learning offers automatic feature extraction and reduces complex signal preprocessing in radar data analysis.

Purpose of the Study:

• To develop an accurate deep learning model for driving behavior recognition using FMCW radar data.
• To address the domain shift challenge in radar-based behavior classification.

Main Methods:

• A Multi-channel Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network (MCT-CNN-LSTM) was developed to capture spatial and temporal features.
• Efficient Channel Attention (ECA) module was integrated for adaptive feature channel weighting.
• Domain-adversarial training was applied to extract common features across domains, mitigating domain shift.

Main Results:

• The proposed MCT-CNN-LSTM model achieved a high accuracy of 97.3% on a real-world measured dataset.
• The integration of multi-scale and channel-time attention modules improved feature representation.
• Domain-adversarial training effectively reduced the discrepancy between source and target domains.

Conclusions:

• The MCT-CNN-LSTM network provides an effective solution for accurate driving behavior recognition from FMCW radar signals.
• The proposed method demonstrates robustness in bridging domain gaps, enhancing generalization capabilities.
• This approach holds significant potential for improving automotive safety systems through advanced driver monitoring.