Reliability of Deep Neural Networks for an End-to-End Imitation Learning-Based Lane Keeping

In recent years, end-to-end imitation learning-based deep neural networks (DNN) have been successfully applied to several autonomous driving tasks. Meanwhile, however, the large uncertainties with respect to the reliability of the DNN’s decision-making mechanism significantly limit the application of DNN-based approaches in the safety-critical domain of autonomous driving. In this work, a comprehensive analytical method of the DNN’s decision-making mechanism is proposed for developing a reliable imitation learning-based end-to-end visual solution for automated lane keeping on the highway in a co-simulation environment. During the mechanism analysis, three progressive and crucial questions covering from the DNN’s decision-making basis to decision itself can be explicitly answered using the modified algorithms of Explainable artificial intelligence (XAI) and the theoretical knowledge of vehicle dynamics; During the model development, the DNN’s architecture is stepwise improved based on the intermediate analytical and test results in the imperfect and failure cases, aimed at increasing the DNN’s lane keeping imitation accuracy and robustness from a superficial point of view as well as enhancing the reliability of the DNN’s decision-making mechanism from an essential point of view. The DNN’s lane keeping performance is quantitatively evaluated based on a proposed evaluation method considering 3 vital perspectives regarding vehicle dynamics. The ultimate test results show that the final proposed DNN model, which has a reliable decision-making mechanism, achieves a satisfactory lane keeping performance with a high imitation accuracy and robustness, and outperforms the state-of-the-art lane keeping approaches based on the end-to-end DNN models and the traditional modular solution.