AI-Powered Safety Management for Construction and Plant Sites

While obligations for industrial safety have been strengthened through measures such as the enforcement of the Serious Accidents Punishment Act, the accident rate at construction sites remains at a high level. Existing personnel-centered monitoring has limitations in managing complex site hazards in real time, making the introduction of an AI-based automatic monitoring system using CCTV footage increasingly urgent.

KOLON BENIT has developed AI Vision Intelligence, which combines multimodal AI with hybrid GPU-NPU infrastructure. This system combines CCTV footage with on-site data to detect the following hazards and provide precise alerts.

• Detects failure to wear hard hats, unauthorized entry into hazardous zones, and absence of a signal person
• Simultaneously determines whether nearby workers and a signal person are present when heavy equipment approaches
• Provides context-based precise alerts, such as "a signal person needs to be deployed immediately"

The system architecture adopts a hybrid structure for efficient workload distribution. GPU handles data training and video processing, while Rebellions' NPU server is dedicated to inference. In addition, the vLLM library was applied to increase inference throughput, and the system was designed via the Rebellions SDK to enable existing GPU-based code to be smoothly migrated to the NPU environment at no additional cost.

This solution, deployed at a KOLON GLOBAL construction site, achieved results of reducing power consumption by up to 50% and heat generation by about 45% compared to existing GPU servers. This leads to lower server room operating costs and an overall reduction in TCO (Total Cost of Ownership). This collaboration is meaningful as a case of building a Sovereign AI package that combines a domestic AI model with a domestic NPU.