Genetic Programming for Channel Selection from Multi-stream Sensor Data with Application on Learning Risky Driving Behaviours

Unsafe driving behaviours can put the driver himself and other people participating in the traffic at risk. Smart-phones with builtin inertial sensors offer a convenient way to passively monitor the driving patterns, from which potentially risky events can be detected. However, it is not trivial to decide which sensor data channel is relevant for the task without domain knowledge, given the growing number of sensors readily available in the phone. Using too many channels can be computationally expensive. Conversely, using too few channels may not provide sufficient information to infer meaningful patterns. We demonstrate Genetic Programming (GP) technique's capability in choosing relevant data channels directly from raw sensor data. We examine three risky driving events, namely harsh acceleration, sudden braking and swerving in the experiment. GP performance on detecting these unsafe driving behaviours is consistently high on different channel combinations that it decides to use.