Explainable AI for Real-Time Object Detection in Autonomous Driving

The increasing complexity of Artificial Intelligence (AI) models, especially in applications like autonomous driving, healthcare, and other systems, requires not only high accuracy but also transparency in the decision-making processes. This paper presents an approach to integrating Explainable AI (XAI) methods with YOLOv8, a state-of-the-art object detection model, to provide interpretability of its predictions in real-time autonomous driving environments. We focus on implementing various Class Activation Mapping (CAM) methods, such as Grad-CAM and HiResCAM, as well as Layer-wise Relevance Propagation (LRP) to create an explanation framework. We also combine CAM and LRP to make use of the strengths of both methods, improving the overall explainability. This method achieves good performance, running object detection at 50 to 60 frames per second (FPS) on a standard dedicated GPU, and around 24 FPS when running alongside the explainability models. Furthermore, this approach maintains a high detection accuracy, due to its post hoc nature, making sure that the XAI implementations do not affect the model’s detection performance. Our experimental results shows that the combined XAI approach not only preserves the high detection accuracy of YOLOv8 but also provides more meaningful visual explanations using the XAI methods, showing that this integrated XAI methods could improve the trustworthiness and safety of AI-driven systems in autonomous vehicles and other possible fields.