Analysis of the computational performance of machine learning models for fatigue detection using facial analysis in an embedded system

Abstract

Fatigue detection using facial analysis with computer vision and machine learning in embedded systems is a non-invasive solution, but it faces challenges in evaluating the computational resources used on hardware devices like a Raspberry Pi. Previous studies have investigated fatigue and drowsiness detection using facial features related to mouth and eye opening and computer vision-based classification models, without considering the computational performance of these solutions. This research aims to evaluate the performance of machine learning models that use facial point features for fatigue detection implemented on a Raspberry Pi, considering metrics such as CPU usage, memory, inference time, and processing speed. To this end, support vector machine, random forest, and decision tree models are evaluated to assess their computational performance. The results show that no single model is superior in all aspects, as decision trees demonstrate better detection of events like yawning (YAWN), while the SLEEP class proves to be the most complex for all models. Furthermore, the decision tree stands out for its lower latency and greater adaptation to real-time applications, so the selection of the model is based on a relationship between accuracy, latency and consumption of processing resources