Pattern Recognition Techniques to Infer Driver Intentions

Pattern Recognition Techniques to Infer Driver Intentions Hiren M. Mandalia Dr. Dario Salvucci, Ph.D. Driving is a complex task that requires constant attention from the mind and the body. Automobile drivers today are under high risks, thanks to the ever-expanding telematics industry, cell-phone driving and other distractions. Inferring driver intentions, especially critical ones like changing lanes, is therefore necessary for any intelligent driver support system. This thesis explores different methods to infer driver's intention to change lanes. Experimental data were collected that observed driver's behavior (e.g. speed, steer angle, gas pedal pressure) and environmental data around the driver (e.g. distance of the car in front). With the hypothesis that such data would display significantly different patterns during a lane change versus lane keeping, this problem was formulated as a pattern recognition problem. Two different techniques were studied in detail to solve this problem, namely support vector machines (SVMs) and hidden markov models (HMMs). These two machine learning techniques showed promising results. SVMs have been particularly effective in early detection of lane changes with a very high sample-bysample prediction rate. In addition, this thesis compares these techniques with a new “mind tracking” approach [9, 10] and also proposes a new graph-based approach to detect lane changes.