RB-XAI: Relevance-Based Explainable AI for Traffic Detection in Autonomous Systems

Recent progress in artificial intelligence has brought about a revolutionary era spanning various sectors such as transportation, healthcare, finance, and cybersecurity. In the field of transportation, these advancements empower autonomous systems to precisely perceive their surroundings, make swift and secure decisions, and operate seamlessly without human intervention. In the case of Autonomous Vehicles (AVs), the perception phase, involving tasks like road surface extraction and on-road object detection, is crucial. While AVs offer benefits like improved road safety, convenience, and efficiency, they grapple with the “black box” challenge, symbolizing the opacity of their decision-making processes to humans. This absence of transparency poses acceptance, regulatory, ethical, and security issues. To tackle these challenges, our work leverages Concept Relevance Propagation (CRP), a bias-resistant Relevance-Based (RB) eXplainable Artificial Intelligence (XAI) algorithm to proffer transparent concept - level explanations to the behavior of object detection models used in AVs, for traffic perception. CRP goes beyond traditional attribution maps, by generating explanations that automatically identifies and visualizes relevant concepts within the input space. This insight sheds light on the crucial areas responsible for the behavior of object detection models used in AVs. This research aims to boost transparency, trust, and understanding in the field of AVs, fostering safer and more widely accepted autonomous systems.