Driver Distraction Detection on Edge Devices via Explainable Artificial Intelligence

Driver Attention Monitoring is a challenging research task, with several complex behavioral distractions to recognize, and the need to use non-invasive on-board systems. Recent advances in deep learning have great potential in this field. This research aims to propose an architectural solution based on a deep convolutional neural network, deployed on an edge device. For this purpose, a publicly available dataset has been exploited to recognize various distracting driver behaviors. The model has been validated using explainable AI techniques. Experimental studies show that the proposed architectural solution, deployed on an NVIDIA Jetson Nano board, achieves a throughput of 11 frames per second and an accuracy of 92%. In contrast to the recent state-of-the-art solutions, the proposed approach covers all the relevant requirements: (i) it is non-invasive; (ii) it has a complexity suitable for popular edge devices; (iii) it allows a reliable validation of the different driver distractions via explainable AI; (iv) it achieves a competitive accuracy.