Driver distraction detection using a perspective-integrated multi-task and multi-expert learning framework

Driver distraction is a leading cause of traffic accidents worldwide, making reliable driver monitoring essential for intelligent transportation systems. However, most deep learning-based approaches rely on a single camera perspective, which reduces robustness under occlusion, a limited field of view, and varying illumination. This study proposes PIMENet, a Perspective-Integrated Multi-Expert Network that formulates multi-perspective driver distraction detection as a perspective-aware expert-routing problem. Instead of applying late fusion to independently trained classifiers, PIMENet introduces a Perspective-Aware Multi-Gate Mixture-of-Experts (PA-MMoE) module that organizes experts according to semantic visual perspectives. The framework defines unimodal experts for body, face, and hand perspectives, as well as pairwise experts for body-face, body-hand, and face-hand interactions. Task-conditioned attention gates dynamically route each perspective-specific task through the most relevant experts, while impartial multi-task learning promotes balanced optimization across tasks. In addition, a parameter-efficient ConvFormer encoder, developed by replacing the final stage of ConvNeXt with PoolFormer blocks, reduces model complexity while improving classification performance. Under a strict driver-independent evaluation protocol, PIMENet achieves 97.01% accuracy on the 10-class 3PDD dataset and up to 89.90% accuracy on the more challenging 22-class 100-Driver dataset, demonstrating highly competitive performance in multi-perspective driver distraction detection. Deployment profiling on an NVIDIA Jetson Xavier platform further indicates that the proposed framework is feasible for decision-level in-vehicle monitoring. These findings suggest that PIMENet can support more reliable driver monitoring systems and contribute to safer intelligent transportation applications.