Towards Real-Time Explainable AI: Using Class Activation Mapping for Brake Prediction in YOLO-Based Systems - Figure 3

In this section, we outline the complete workflow of the proposed framework as shown in Figure 3, detailing each stage from data acquisition to evaluation of the trained model and explainability layer. Firstly, the process begins with data collection, where a diverse dataset is gathered to support robust model training. After this, data preprocessing steps such as imageresizing and normalisation are applied. Following this, the model training phase employs YOLOv8 with transfer learning to achieve real-time and highly accurate detections. A performance analysis is conducted using key statistical metrics to evaluate the model’s effectiveness. Furthermore, to enhance transparency and interpretability, XAI techniques are incorporated in the YOLOv8 architecture, allowing for visualisation of model decisions through heatmap generation. These visual insights provide a deeper understanding of how the model identifies key features, ensuring its reliability in practical applications.