Comparative Analysis of Two Approaches for Improved Site Safety: Domain Approach versus General End-to-End Compliance Reasoning

Field compliance checking seeks to automate the process of extracting safety requirements, capturing site information, identifying violations, and generating documentation for different responsible parties to improve construction site safety. However, existing safety checking methods are typically limited in their ability to employ domain knowledge and multimodal data, and they often lack generalizability and scalability. There is also limited effort to automatically generate natural-language descriptions of compliance-checking results for documentation purposes. Aimed toward addressing these gaps, this paper studies and compares two approaches for automated safety checking: a domain approach and an end-to-end approach. The domain approach integrates existing domain-specific information extraction, site visual information detection, and knowledge graph–based compliance reasoning methods. The end-to-end approach relies on multimodal foundation models. The comparative analysis evaluates the information extraction and compliance reasoning capabilities of both approaches in terms of performance, errors, and limitations. The results showed that the domain approach achieved strong performance with good potential for generalizability and adaptability. In contrast, the end-to-end approach requires additional fine-tuning to achieve comparable results.