Introduction: The aspects of safety and reliability are inherent in autonomous driving systems, particularly in predicting and evaluating the risks of failures before they lead to unsafe conditions. The objective of the study is to enhance failure prediction by diversifying data learning to enhance the steering angle prediction. The main goal of the solution is to develop a scalable model. Methods: The proposed framework uses a convolutional neural network trained with a pilot-based approach, where saliency maps are generated via Visual Backpropagation. The model is initially trained on the Udacity Jungle simulated dataset and subsequently extended with the Udacity Lake dataset and a real-world driving dataset to increase the environmental variety. The steering angles of all the datasets are normalised through min-max scaling to make them have the same numerical range and learning behaviour. results: The saliency-enhanced CNN demonstrated improved interpretability and prediction reliability by achieving higher pixel accuracy and lane boundary precision than the baseline CNN. Enhanced feature localization was also confirmed by the integration of VisualBackProp, which resulted in enhanced SSIM scores. Results and Discussion: The results are significant in maintaining the value of various datasets, and the uniform preprocessing improves the model's robustness under various driving conditions. Salient map learning provides interpretable indications of steering behaviour, enabling the prediction of failures in alternative ways. Conclusion: The proposed framework provides better failure prediction and enhanced generalisation, which is why it is applicable to real-world scenarios such as autonomous driving.
Dangi et al. (Fri,) studied this question.