Improving 3D Object Detection for Autonomous Driving – A Case Study of Data-Driven Development

Autonomous Driving (AD) solutions are poised to revolutionize mobility, driving significant R&D efforts. However, scaling this technology presents major challenges, necessitating a re-evaluation of automotive R&D processes. Agile and DevOps methodologies are crucial for faster innovation cycles and meeting the demand for software-defined features. Artificial Intelligence (AI), particularly Machine Learning (ML), is integral to advancing AD systems, requiring a shift towards "data-driven development." While AI models offer robustness, their statistical behavior poses risks, which are addressed by current standards like ISO 21448 “Safety of the intended functionality (SOTIF)” and ISO 8800 “Road Vehicles – Safety and artificial intelligence”, currently under development. Real-life AD applications demand adaptation to new data, requiring advanced techniques like Active Learning and Continuous Learning. AD systems, as software-defined products, require constant updates and integration of AI components, following automotive industry standards. Drawing from MLOps/AIOps practices, our study applies data-driven development to a camera-based 3D Object Detection case study, iteratively improving the detector through a combination of Active Learning and Semi-Supervised Learning, demonstrating the transformative potential of a well-implemented “Data Loop” in the automotive industry.