Embedded TinyML for Driver Drowsiness Detection

Abstract

Driver drowsiness is a relevant road-safety risk. This paper presents the deployment of a TinyML image classifier on the Seeed Studio XIAO ESP32-S3 Sense. A convolutional neural network was trained on the Driver Drowsiness Dataset using Edge Impulse, quantized to INT8 and integrated into an Arduino firmware. Float32 and INT8 models achieved 99.9485% accuracy on the adopted image-level test partition. The run_classifier() function required an average of 164.4 ms on the microcontroller, corresponding to approximately 6.1 inferences per second. A heuristic accumulated-score mechanism was implemented to integrate consecutive predictions. Because the dataset contains facial frames extracted and cropped from videos and was partitioned at the image level, the results do not represent subject-independent or real-driving performance. The experiment demonstrates the computational feasibility of executing the quantized classifier on a compact embedded device.