Driver assistance and performance monitoring systems are currently being applied in modern cars in order to enhance safety. However, these systems have to answer certain concerns raised by manufacturers, legislators and users. These include, degree of intrusiveness (warning messages, tactile feedback, taking control of the car), ability to respond to different driving contexts and system reliability under varying road and environmental conditions and driver reliability. By combining inexpensive and non-intrusive sensors with state-ofthe- art signal processing, probabilistic theory and artificial intelligence for signal analysis and modelling, it is possible to present a solution to all the above concerns to a certain extent. To investigate this extent, highway scenario simulator experiments have been conducted including 30 drivers in normal physical condition and impaired conditions due to lack of sleep. A simulator equipped with a near-infrared eye-gaze tracker, strain gauges to measure force on the steering wheel column (SWC), and potentiometers to measure steering wheel and throttle angle has been used. In addition to these core sensors, two webcams have been implemented to view the driver and to track lane-keeping. Raw data have been obtained comprising eye movement, force on SWC, vehicle speed, lane deviation, and human activity from the webcam. The data are first processed up to a level where all signals are one dimensional and continuous. Secondly, metrics have been derived using derivatives, histograms and entropies of the signals. These metrics are then tested against a ground truth risk level obtained from a driver survey and from independent observers. After selecting the best metrics for driver performance indication, different time windows for metric derivation are compared and the driver sessions are classified by a Fuzzy Inference System The system works well on the simulator data, with a 98% correct classification rate and is now being implemented in real conditions on real roads.