Distracted Driving Behavior Recognition Based on Dual-Stream CNN and Attention Mechanism

Distracted driving remains a major contributor to traffic accidents, emphasizing the need for robust behavior recognition systems capable of operating under complex and dynamic driving conditions. This study proposes a comprehensive framework that integrates a Dual-Stream Attention-Driven Convolutional Network (DACNet) with an Adaptive Attention Integration Strategy to enhance the recognition of unsafe driver behaviors. DACNet employs a spatial stream to capture static visual cues from individual frames and a temporal stream to model motion patterns across consecutive frames, enabling the network to learn rich spatial–temporal representations. By incorporating an attention mechanism, the model dynamically emphasizes salient regions and time segments, ensuring that the most behavior-relevant information is prioritized during feature extraction. The Adaptive Attention Integration Strategy further refines this process by adjusting attention weights based on contextual relevance, mitigating the influence of environmental noise and variations in driver actions. Through cross-stream attention, the strategy strengthens the interaction between spatial and temporal features, producing a more discriminative multimodal representation. Regularization is adopted to stabilize attention distribution and prevent overemphasis on isolated features, resulting in improved robustness across diverse scenarios. Extensive experiments on multiple driving-related datasets demonstrate that the proposed framework achieves superior accuracy, precision, and robustness compared with existing state-of-the-art methods. The results validate the effectiveness of combining dual-stream architectures with adaptive attention mechanisms for distracted driving behavior recognition. The proposed model shows promising potential for real-world intelligent transportation systems by offering a reliable and interpretable approach to identifying distraction-related risks.