Optimizing Driver Monitoring System Performance on TI-TDA4VM Edge Device

This research focuses on developing an advanced Driver Monitoring System (DMS) designed to address critical issues of driver fatigue and distraction, both major factors in road accidents. The system employs facial landmark detection and computer vision techniques to classify various driver states, including alertness, sleepiness, distraction, and drowsiness. It incorporates essential features like eye closure detection, gaze tracking, yawn detection, head pose estimation, and gender and age estimation. These features are captured using the IMX219 camera sensor with infrared LEDs, ensuring consistent and reliable performance in diverse lighting conditions. To enable real-time processing, the system is built on the Texas Instruments TDA4VM platform, taking advantage of its processing capabilities and deep-learning accelerators, with BeagleBone AI-64 serving as the development board. The integration of the Open Neural Network Exchange (ONNX) and Texas Instruments Deep Learning (TIDL) frameworks allows the system to operate efficiently on the edge, eliminating the need for external computing devices. Experimental evaluations indicate the system's high effectiveness, achieving approximately 90% accuracy and a processing speed of 22 frames per second (FPS). While various existing DMS solutions face challenges in balancing real-time performance, accuracy, and computational efficiency, especially under different lighting conditions, our approach successfully meets all these criteria, providing robust and reliable results. These advantages underscore its potential to significantly enhance road safety through more effective driver monitoring.