Driving Style Analysis Using Recurrent Neural Networks with LSTM Cells

摘要

Many publications work on optimization of driving styles in motor vehicles. Most conclude that they can improve energy efficiency through training. In recent years the tools to address those problems evolved towards machine learning. To get appropriate data for learning algorithms we developed a method to judge a driving style with respect to energy efficiency. This approach leveraged handpicked criteria like acceleration extracted from GPS. Like related works, this method does not scale, since it requires substantial preprocessing. The goal of this evaluation was to reduce the resistance energy of a driven trip, while maintaining a natural traffic flow. This was accomplished by mimicking a low-pass filter on the speed profile. On top excessive speeding gets punished. It was possible to use our data with over 1 million kilometers for training a Recurrent Neural Network. In respect to the RNN the training data was used, to let it map the obtained function. The provided data was adjusted in different stages, until it was only the raw GPS data. The RNN learned to handle most GPS errors, only in initial phases the results are mixed. A RNN Network is well suited to handle GPS data and learn higher level features on its own. The result is a NN which judges the driving style using only raw GPS data.