Multisensor Data Fusion For Driver Safety

Motor vehicle crashes are a leading cause of death in the world. Although Millions ofdollars are devoted by Automakers to developing advanced system that can improve the roadsafety and save huge economic and lives losses, the number of road fatalities is still increasing.According to the World Health Organization WHO an estimate of1.3million people werekilled in road injuries in2011compared to1million road fatalities in2000. The US NationalHighway Traffic Safety Administration NHTA reported33,561road crash fatalities in2012compared to32,479in2011. Moreover, road crashes has a destructive economical effect.Among the main reasons of road fatalities, are drowsiness and distraction. A Harvard riskanalysis study estimated the annual cost of crashes caused by cellphone use to be$43billion.Drowsy driving related accident account for$48billion in medical costs each year. Thissituation presents a real challenge for automakers and researchers who need to consecratemore effort to bring state of the art technology to save millions of injuries and huge economicloses. Hereof in this thesis we aim to develop an advanced system able to enhance roadsafety. Some studies estimate that human error accounts for90%of road accidents; thereforewe focused our system on monitoring the driver’s state of alertness. Drowsiness anddistraction was addressed by this system, the detection in real time is done by mean of varioussensors, computer vision and machine learning.Various kind of distraction such as talking on the phone and not looking ahead are aheadbased on some simple observations of the hands and head pose. For drowsiness detectionthe system uses several sensors to acquire input information. The information could beorganized in three categories: information obtained from measuring driving performancemetric such as pressure on accelerating pedal and steering angle, information obtained frommonitoring driver’s visual behavior such as yawning and eye closure. And informationobtained by measuring biomedical signal of the driver such as heart beats. Each of theseinformation was first analyzed separately to extract key features necessary for the calculationof indicators like heart rate variability HRV and percentage of eye closure PERCLOS.Because studies have shown that every single indicator has its own limitation, a combination of several indicators results certainly on much accurate and reliable indicator of the driver’slevel of vigilance. Combining various data of different sensors is known as data fusion. In oursystem we propose a model of data fusion based on Bayesian network. Experiments in vehicleunder real conditions and on laboratory have shown that the system is effective in detectingdrowsiness based on observations of head pose, eye state and mouth. Distraction is alsodetected with good accuracy based of head movement. Finally this paper propose a model fordata fusion.