Summary

  • Rising accident rates and stricter safety regulations are driving demand for AI-powered intelligent monitoring in construction sites
  • Implemented an intelligent monitoring solution using NPU servers and a multimodal AI model pipeline
  • Built a GPU–NPU hybrid system, ensuring high compatibility with existing GPU-based code and AI serving software, enabling easy migration without additional costs or resources
  • Reduced power consumption and heat output, proving the ability to lower monitoring center server TCO

Challenge 

Following the introduction of Korea’s Serious Accidents Punishment Act, industrial safety obligations have become stricter. Yet the fatal accident rate at domestic construction sites remains far higher than the OECD average. This creates an urgent need for advanced accident-prevention systems, not only to protect workers’ lives but also to reduce financial and social costs for companies and society.

Construction sites are complex environments where large numbers of workers and heavy equipment move in close quarters. Relying solely on supervisors and CCTV operators to manage safety risks has clear limitations. This is why AI-powered monitoring systems, particularly those based on CCTV video, are gaining traction as preventative and automated solutions.

However, deploying AI safety monitoring at scale requires solving multiple challenges. Specialized AI models must be developed for construction environments, which means analyzing site-specific hazards and collecting and refining relevant data. These models must then be integrated into a service pipeline that can accurately detect meaningful events and present clear alerts to monitoring staff. Finally, the infrastructure must be deployed inside monitoring center server rooms, where space, cooling, and power supply capacity are all constrained. Traditional GPU-based systems, with their high cost, power demands, and heat generation, impose a significant burden that only grows as AI systems become more advanced.

Solution 

Kolon Benit overcame these limitations by developing AI Vision Intelligence, a differentiated safety monitoring system powered by multimodal AI and hybrid GPU–NPU infrastructure.

At its core, AI Vision Intelligence combines CCTV video with construction site data in a multimodal model architecture. It detects safety risks such as workers not wearing helmets, unauthorized entry into hazardous zones, absence of signalers, or the approach of heavy equipment, and then translates these detections into actionable alerts. Importantly, by leveraging multimodal modeling, the system goes beyond simple object detection to understand the context of risk factors and guide monitoring staff with precise, meaningful event notifications. For example, when heavy equipment is detected on site, the system simultaneously checks for nearby workers and whether a signaler is present. It can then prompt operators to dispatch a signaler immediately and secure the area around the equipment, offering actionable guidance that significantly improves accident prevention. This contextual awareness reduces false positives, ensures accurate detection, and delivers highly specific alerts that enhance preventive measures.

The system architecture is also designed for operational efficiency. Training and video processing are carried out on GPUs, while inference for object detection and language generation runs on Rebellions’ low-power, high-performance NPU servers. This hybrid setup maximizes resource efficiency. For large language models, Kolon Benit applied the high-performance vLLM library, improving both inference throughput and operational convenience. The Rebellions SDK, with its strong compatibility, enabled seamless migration of existing GPU-based code and infrastructure to the NPU environment without incurring additional transition costs or resource overhead.

Result 

Kolon Benit deployed AI Vision Intelligence in the monitoring center of Kolon Global construction sites. By introducing an automated system that identifies risks in advance and communicates them with clarity, monitoring staff can now understand hazards more accurately and respond with preventive actions. This is expected to contribute significantly to reducing accident rates.

The Rebellions NPU servers proved particularly effective in optimizing infrastructure efficiency. Compared to GPU servers, NPU servers cut power consumption by up to 50 percent and reduced heat generation by around 45 percent. These savings translate into lower infrastructure operating costs and, as systems scale, further reductions in TCO across power supply capacity, cooling systems, and overall monitoring center operations.

Beyond performance and cost benefits, this solution represents a milestone for sovereign AI in Korea. Built with Kolon Benit’s own technology, domestic AI models, and Rebellions’ domestically developed NPUs, it demonstrates the creation of a Korean sovereign AI package. Going forward, Kolon Benit aims to expand this ecosystem in line with national AI strategies, delivering practical value to industrial sites while driving the growth of new AI businesses.

“Rebellions’ ATOM is an NPU with an exceptionally wide range of applications. By building a hybrid GPU–NPU system and multimodal inference pipeline, we successfully developed the AI Vision Intelligence solution for safety monitoring. The results confirmed excellent inference performance, energy efficiency, and software compatibility. Based on this success, we will continue to expand a sovereign AI ecosystem built on our own technology.”

— Kolon Benit

Appendix: NPU User Guide

Get Started Now : Prepare a pre-trained model → Compile with the RBLN Compiler → Load the compiled model and run inference