In the rapid development of artificial intelligence(AI), explainability(XAI) in particular plays a critical role in ensuring the reliability and safety of AI systems. Existing XAI studies mainly focus on interpreting the process by which a model derives a specific conclusion, which is essential for providing transparency in AI decision-making in high-risk applications. However, most studies focus on how AI makes correct predictions, and there is a relative lack of research on how to analyze and effectively correct the root causes of errors that occur during the data collection process. Errors that occur during the data collection process can lead to biased model learning, misclassification, and decreased overall system reliability. In general, AI pipelines regard such errors as simple noise and respond by applying outlier detection or noise removal techniques to refine the data, but these approaches have limitations in resolving the root causes of errors. In this study, we propose a feedback-based error correction framework that goes beyond simply detecting data errors to explain why the errors occurred and quickly correct them through a more systematic approach. The proposed framework detects errors early in the data collection phase, classifies the types of errors, identifies the error occurrence mechanism through question-answering($\mathbf{Q A}$)-based cause analysis, and constructs a feedback loop to correct errors through automated correction or human intervention based on the results, thereby continuously improving data quality. In particular, it focuses on data error management in high-risk environments such as autonomous driving, and proposes an environment-adaptive data error detection and correction technique.

The revolutionary development of artificial intelligence is making significant advancements such as Gen AI, and safe autonomous driving technology systems are being researched to be applied to autonomous driving environments that require the application of advanced and reliable technologies. In order to establish a stable autonomous driving environment, object detection, motion prediction are being researched. Currently, road marking recognition is performed in the object detection field to acquire road driving environment information, however, the inference of possible situations and proactive response from road markings are not being conducted. To overcome the limitations, this paper proposes a Proactive Scenario-driven Adaptable Defensive Driving (PS-ADD) Architecture to infer Environment Adaptive Scenario from proactive scenarios and derive Environment Adaptive Preemptive Plan. To infer proactive scenarios, road markings and real-time object information is combined to generate Proactive Driving Approaches.Then, Environment Adaptive Defensive Decision Making (EADDM) algorithm enables Adaptable Defensive Decision Making by deriving realtime environmentally adaptive scenarios and preparing adaptive preemptive plans. PS-ADD Architecture enables to predict and handle potential scenarios in advance in complex driving situations, allowing for stable defensive driving in unanticipated situations.

Autonomous vehicles make decisions and controls based on various object recognition results. The driving environment is characterized by the coexistence of a multitude of objects of varying shapes and sizes. Therefore, the ability to accurately recognise fine-grained objects is essential for accurate object recognition in a variety of changing situations. For object detection, the autonomous vehicle performs bounding box and segmentation to provide the detected object information. However, bounding box and segmentation based object detection has difficulties in identifying objects with complex shapes, small or distant objects, and it is hard to distinguish and detect objects with similar colors to the background or similar colors and textures to surrounding objects. This has limitations for reliable object identification in autonomous driving environments containing a variety of objects, which is a challenge for clear criteria-based object avoidance and collision protection. To overcome these limitations, this paper proposes Edge-Adaptive Depth Estimation(EADE). EADE, the combination of edge extraction and depth estimation, enables detailed edge extraction and partial distance estimation of objects even in environments where object shape and size, surrounding objects, and backgrounds make it difficult to recognise distinct objects, which allows for reliable autonomous decision-making and control based on detailed object collision and avoidance criteria. To validate EADE, experiments were conducted with real-world driving environment image data. The results of EADE demonstrate that detailed object recognition is possible with clear edge recognition and estimation of object distance, even for complex shaped objects such as trees with branches in multiple directions, distant objects, and objects that are difficult to distinguish from the background such as curbs.

The Advanced and radical development of IT technology and artificial intelligence technology have made it possible to develop advanced services Digital Twin, Metaverse, Metatwin-verse, etc using Artificial Intelligence(AI). The results induced from AI present the correct solution when AI performs accurate study and analysis. Specifically, real situations reflecting complex relationships between objects, results from real situations have to be adaptive to convergence situations and then it should be possible to draw conclusions and make decisions that are not limited to specific situations. So, it is essential to conduct AI based study and analysis by considering these real world characteristics to provide digital twin services based on metaverse. Recently, there are many studies on Graph Neural Network(GNN) and services applied to GNN for learning the relationship between objects detected in real situations. Accordingly, this paper proposes a metaverse-based Digital Twin Convergence Service(DTCS) including spatial elements strategy that is possible to draw accurate conclusions in a changing convergence situation. DTCS is able to conduct causal reasoning and association learning between objects considering directions and distances change characteristics between objects and this is possible to make correct solution and decision making in the process of simulation and analysis of digital twin. In that DTCS proceeds by considering distance and changing angle between objects, this overcomes the limitation of existing GNN which only considers the degree of association or assigns the same parameters to connected objects. DTCS would be possible to expand the advanced services of Metatwinverse in that it is possible to have robust learning based conclusions in real-time changing convergence situations.