This presentation introduces a paradigm shift in automotive ECU testing: replacing brittle, rule-based image processing with Deep Learning. Currently, automating the visual validation of vehicle outputs (e.g., lights, displays) relies on classical methods like thresholding. These frequently fail under industrial glare, shadows, or occlusions, causing false test failures.
We present an AI-based perception pipeline utilizing YOLOv8 for End-to-End verification. Through rigorous comparative benchmarking against five traditional computer vision techniques, we demonstrate that YOLOv8 completely overcomes these environmental vulnerabilities. Crucially, it achieves this robust accuracy with an inference latency of under 25ms, ensuring real-time Hardware-in-the-Loop (HiL) compatibility. Integrated via ROS2/MQTT for seamless CAN/LIN signal fusion, this deep learning approach eliminates manual inspection bottlenecks. Ultimately, this transition from thresholds to learned representations enables highly scalable, robust, and 24/7 automated safety-critical testing for Volkswagen.
