NN-based measurements for driving pattern classification

Abstract

This paper aims at showing how to classify driving patterns in terms of primitives such as acceleration, deceleration and turning, using neural networks. In particular a multilayer perceptron with back-propagation learning algorithm is used. The considered feature space is reduced to a very restricted couple of sensors: accelerometer and GPS receiver, which characterize many commercial low-cost inertial navigation systems (INS). Sensor-driven input patterns are used for classification over output driving primitives. The ease of this approach holds true since GPS data coupled with forward and lateral accelerations are sufficient for describing much of the semantics of driving scenarios. This argument is supported by real observations on different types of vehicles and different types of drivers. These measurements show that, in normal conditions, the road geometry implies a vehicle to adopt a well-defined behaviour, which can therefore straightforwardly be characterized.