Quantified LIME explainability and class-specific dual-branch trust calibration for safety-critical road-scene perception in autonomous navigation

The current research focuses on the concept of safety-critical autonomous navigation, where the perception system not only needs to be accurate but also transparent, class-specific, and robust in the face of uncertainty. In this paper, the concept of RoadXAI-Trust, an explainable dual-branch perception system, has been introduced, which combines semantic segmentation, object detection, and trust-based recommendation. In the semantic segmentation branch, the performance of the models DeepLabV3+, UNet++, and TransUNet + + has been evaluated on the CamVid dataset. Recent evaluations have confirmed that the best performing model is the DeepLabV3 + model, which obtains validation set metrics of 0.5973 mIoU, 0.7183 Dice, and 0.9158 pixel accuracy, along with their respective test set metrics of 0.5252 mIoU, 0.6428 Dice, and 0.8909 pixel accuracy. In the object-detection branch, the YOLOv8 model was trained on an eight-class vehicle dataset, which performed well, obtaining precision of 0.9746, recall of 0.9512, mAP@50 of 0.9856To make the perception system transparent, the concept of LIME has been introduced, where the quality of class-specific explanation is measured based on positive coverage, mean contribution weight, and stability for Road, Car, and Pedestrian. These values are then combined with the statistics of the object detection confidence and passed to the trust core, which provides the decision as High-Trust, Moderate, or Caution/Review. From the experimental results, it can be concluded that the current research provides robust and transparent perception of the roadway and vehicles, while also following a precautionary principle for pedestrian perception.