Driver behavior recognition under multiple illumination conditions based on attention mechanisms

This study presents a dual-stream spatiotemporal attention network (DSTA-Net) for robust driver behavior recognition under complex illumination conditions. Based on the SlowFast backbone, DSTA-Net incorporates the Temporal-Channel Attention Module (TCAM), Temporal Attention Focusing Algorithm (TAFA), Adaptive Rank Pooling Dynamic Image Generation (ARPDIG), and Coordinate Attention (CA) mechanism to enhance temporal sensitivity and illumination adaptability. A computationally efficient (2+1)D convolutional structure is adopted to reduce computational cost while maintaining spatiotemporal representation capability. Evaluated on the proposed MAID-Behav dataset with multi-view and multi-illumination scenarios, DSTA-Net achieves 98.76% accuracy under normal lighting and shows improved performance under low-light environments, outperforming state-of-the-art models by over 7%. The proposed model provides a spatiotemporal modeling framework that may be applicable to intelligent driving scenarios under the evaluated experimental settings.