A High Performance and Robust FPGA Implementation of a Driver State Monitoring Application.

Area of Science:

• Computer Vision and Machine Learning
• Automotive Safety Systems
• Reconfigurable Hardware Acceleration

Background:

• Driver drowsiness is a significant cause of road accidents.
• Existing Driver State Monitoring (DSM) systems often face challenges with real-time processing and accuracy.
• Machine learning methods, like Ensemble of Regression Trees (ERTs), show promise for facial landmark analysis.

Purpose of the Study:

• To develop and implement a high-performance Driver State Monitoring (DSM) application for detecting driver drowsiness.
• To accelerate the facial landmark alignment process using reconfigurable hardware.
• To enhance the robustness and accuracy of drowsiness detection by incorporating facial shape coherency rules.

Main Methods:

• Utilized the Ensemble of Regression Trees (ERTs) machine learning algorithm for aligning 68 facial landmarks.
• Ported and adapted an open-source ERTs implementation for reconfigurable hardware to accelerate frame processing speed.
• Employed unconventional hardware acceleration techniques to manage large data transfers and low data reuse in computational kernels.

Main Results:

• Achieved a high frame processing rate of up to 65 frames per second.
• Demonstrated high robustness and accuracy by ignoring false detections and estimations.
• Reported sensitivity and precision in yawning recognition reaching 93% and 97%, respectively.

Conclusions:

• The implemented DSM algorithm on reconfigurable hardware provides a high-performance solution for driver drowsiness detection.
• The developed hardware acceleration techniques are effective for machine learning applications with demanding data transfer requirements.
• This approach offers a promising direction for improving automotive safety through advanced driver monitoring.