Professor Choi Jung-wook’s research team at AI Hardware and Algorithms Lab from the School of Electronic Engineering at Hanyang University has developed a high-performance AI application optimization technology. This technology utilizes domestic NPUs(Neural Processing Units) to provide real-time walking assistance for the visually impaired.

The research focused on driving large Vision-Language Models(VLM) on low-power NPUs in real-time to immediately guide visually impaired users through surrounding hazards.

Existing VLM-based assistive technologies offer excellent situational awareness but require high-power GPUs due to massive computational demands. This has posed significant challenges for wearable or mobile devices due to battery life and overheating. Furthermore, most existing systems are passive(QA-only), meaning they only respond when a user asks a question, making it difficult to react to sudden, unexpected situations.

To solve this, Professor Choi’s team introduced Rebellion’s ‘ATOM™’, a domestically produced NPU, and proposed two core optimization technologies specialized for it. Firstly, ‘Frame Selection & Batching’ processes continuous video input efficiently by aligning it with the NPU's fixed computational structure. Next, ‘2-Stage Prompting’ describes the situation in the first stage and filters out unnecessary information for walking in the second stage. This allows the AI to actively detect risks(Reminder) and provide concise guidance.

Experimental results showed that this technology reduced End-to-End Latency from the 14-second range to the 4-second range—a 69.1% improvement—securing real-time performance. In demo tests using actual urban driving footage, the system proved its capabilities by independently recognizing hazards, such as an approaching motorcycle, and providing voice warnings. In recognition of these technical achievements, the project received the Excellence Award at the ‘Capstone Design Fair’ hosted by the Seoul RISE Industry-University-Local Cooperation Foundation.

This is highly significant as it opens the door for commercialization by shifting AI systems from power-hungry GPUs to low-power, high-efficiency domestic NPU environments. Based on this secured 토토사이트 악성코드, the research team plans to expand their research into various On-device AI fields in the future.

This study was conducted as part of the ‘2025 High-Performance Computing Support’ project managed by the National IT Industry Promotion Agency(NIPA), receiving support for NPU-based high-performance computing resources.