Convolutional Dual-Attention-Network (CDAN): A multiple light intensities based driver emotion recognition method

Abstract

Driver emotion recognition is critical for enhancing traffic safety and influencing driver behavior. However, current methods struggle to accurately classify emotions under variable lighting conditions such as bright sunlight, shadows, and low light environments, resulting in inconsistent feature extraction and reduced accuracy. Moreover, many approaches incur high computational costs and excessive feature exchanges, limiting real-world deployment in resource-constrained settings. To address these challenges, we propose the Convolutional Dual-Attention Network (CDAN), a novel framework designed to mitigate the impact of light intensity variations in driving scenarios. Our framework integrates Multi-Convolutional Linear Layer Attention (MCLLA), which leverages linear attention augmented with Rotary Positional Encoding (RoPE) and Locally Enhanced Positional Encoding (LePE) to capture global and local spatial relationships. Additionally, a Convolutional Attention Module (CAM) refines feature maps to improve representation quality. Evaluations of MLI-DER, modified KMU-FED, and CK+ datasets demonstrate its enhanced effectiveness compared to existing methods in handling diverse lighting conditions.