This thesis presents an LLM-based approach to requirements engineering for ADAS cameras, aiming to generate system requirements by utilizing new customer requirements and existing reference projects. This is motivated by the fact that requirements engineering is a critical yet highly manual and time-consuming process, especially in complex domains such as ADAS. Three different architectures are introduced and evaluated: a baseline LLM architecture, a RAG architecture and a fine-tuned LLM architecture. Evaluation is conducted using both automated metrics (BLEU, ROUGE, BERTScore and sentence cosine similarity) and expert reviews that evaluate the quality, correctness and feasibility of the requirements.

Results show that RAG and fine-tuning significantly improves lexical and semantic performance, outperforming the baseline across all metrics. Expert reviews highlight that overall the generated requirements have good quality, correctness and feasibility. Some limitations still remain to fully automate the generation process, where requirement engineers are needed to review and approve the generated requirements. The approach still demonstrates potential to support and accelerate requirements engineering workflows.

This work delivers a custom LLM-based tool for ADAS requirement generation, insights into the effect that RAG and fine-tuning techniques have and practical recommendations for integrating such tools into industry processes. Future work is needed to ensure that both RAG and fine-tuning techniques utilize high-quality requirement data, with requirement engineers approving inputs across diverse projects and authors.