In this paper, we explore different Convolutional Neural Network (CNN) architectures to extract features in a Time to Lane Change (TTLC) classification problem for highway driving functions. These networks are trained using the HighD dataset, a public dataset of realistic driving on German highways. The investigated CNNs achieve approximately the same test accuracy which, at first glance, seems to suggest that all of the algorithms extract features of equal quality. We argue however that the test accuracy alone is not sufficient to validate the features which the algorithms extract. As a form of validation, we propose a two pronged approach to confirm the quality of the extracted features. In the first stage, we apply a clustering algorithm on the features and investigate how logical the feature clusters are with respect to both an external clustering validation measure and with respect to expert knowledge. In the second stage, we use a state-of-theart dimensionality reduction technique to visually support the findings of the first stage of validation. In the end, our analysis suggests that the different CNNs, which have approximately equal accuracies, extract features of different quality. This may lead a user to choose one of the CNN architectures over the others.     «

In this paper, we explore different Convolutional Neural Network (CNN) architectures to extract features in a Time to Lane Change (TTLC) classification problem for highway driving functions. These networks are trained using the HighD dataset, a public dataset of realistic driving on German highways. The investigated CNNs achieve approximately the same test accuracy which, at first glance, seems to suggest that all of the algorithms extract features of equal quality. We argue however that the tes...     »