ActiveDrive – Driver Drowsiness Detection System

Abstract

Drowsiness and fatigue behind the wheel pose a serious traffic safety concern, as it severely limits a person’s cognitive and motor functions, making the driver unable to put enough attention on the road and perform the actions necessary for a safe journey. Ever since the dawn of the automobile, drowsy and fatigued drivers have posed a serious threat, not only to themselves but to other bystanders and fellow drivers. Moreover, driving after more than 20 hours without sleep is the same as driving with a blood-alcohol content of 0.08 percent, which is the legal limit in the United States. Unfortunately, pinpointing the exact number of drowsy-driving collisions, injuries, and fatalities is still impossible. Crash investigators can search for signs that sleepiness had a role in a collision, but these signs aren't always obvious or clear. Because indications of exhaustion are difficult to detect, a motorist may be unaware that he or she is weary. Micro-sleep — brief, involuntary bouts of inattention – may also occur in certain persons. At highway speed, the car will traverse the length of a football field in the 4 or 5 seconds a driver experiences micro-sleep. Drowsy driving is involved in roughly 100,000 police-reported collisions each year, according to the National Highway Traffic Safety Administration. More than 1,550 people have died and 71,000 have been injured as a consequence of these collisions. However, because it is difficult to verify whether a motorist was sleepy at the time of a collision, the true figure might be much higher. Our paper proposes a fast and accurate solution to detect driver fatigue and sleepiness without the use of heavy equipment for e.g., breath analyzers, our approach consists of a compactsized and encased Nvidia Jetson Nano, that is running a MobileNet CNN architecture, that has been pretrained on thousands of subject images, retrieved from various datasets such as NTHUDDD, UTA-RLDD, DROZY etc. Our model has achieved a validation accuracy of 97.0%, compared to other models such as AlexNet and ResNet50, which achieved an accuracy of 93.2% and 95.8%, on the NTHU-DDD dataset benchmark