Towards Real-Time Explainable AI: Using Class Activation Mapping for Brake Prediction in YOLO-Based Systems

Abstract

Predicting brake light activation in semi-autonomous vehicles (SAVs) is critical for improving road safety and optimising adaptive cruise control (ACC) systems. Despite the remarkable performance of deep learning-based object detectors such as YOLOv8, their inherent opacity in decision-making processes limits transparency, interpretability, and user trust factors essential for the deployment of safety-critical systems. This study introduces a novel, explainable framework that integrates state-of-the-art attribution-based explainability techniques, including EigenCAM, EigenGrad-CAM, LayerCAM, and HiResCAM, into the YOLOv8 architecture. The proposed framework systematically analyses activation patterns within the model, generating fine-grained saliency maps that highlight the spatial regions most influential in brake light detection. Using a real-world vehicle dataset comprising diverse lighting and environmental conditions, the model is fine-tuned to predict binary brake light states (car_BrakeOn and car_BrakeOff). Comparative experiments demonstrate the ability of these CAM-based methods to provide interpretable visual explanations while maintaining detection accuracy. This study is the first to explore the integration of explainability techniques within YOLO-based brake light detection systems for semi-autonomous vehicles (SAE Levels 2 and 3), addressing a critical gap in the literature. By bridging the divide between black-box AI models and human-understandable reasoning, this research promotes transparency, accountability, and user trust in AI-driven perception systems for intelligent transportation. The findings contribute a foundational step toward the broader adoption of explainable and reliable AI in safety-critical SAV applications.