Prediction of lane change by echo state networks

Abstract

Lane change prediction can reduce traffic accidents and improve traffic flow. To predict lane changes variables which describe lane changes are needed. Recent studies used different classifiers and different inputs for lane change classification and prediction. Here, different methods are used to extract the relevant input variables from a data set which was generated from a naturalistic driving study in the urban area of Chemnitz, Germany. First variables which show different characteristics for left and no lane changes were chosen. The variables contained driver attributes (for instance gazes), environment attributes (for instance distance to other vehicles) and vehicle attributes (for instance velocity). Second, different combinations of these input variables were analyzed with the principal component analysis. In the end, the best combinations were used to classify left lane changes with an Echo State Network and a feedforward neural network. The Echo State Network achieved high area under the curve values, true positive rates and low false positive rates for the classification with a majority of the input combinations. The feedforward neural network predictions were inferior of those to the Echo State Network.