국민대 김장호 교수 연구팀, 자율주행 3D 객체 탐지 AI 경량화 기술 개발

National University of Korea (NUK) professor 김장호's research team led by graduate students 조현준 and 안상호 has developed a lightweight technology called 'SharedKD' for autonomous vehicles, as part of a joint research effort with modern automobile. This technology, presented at the 2026 Design Automation Conference (DAC), effectively reduces the size of 3D object detection models used in autonomous driving by utilizing a single model with both the teacher and student components.

The key feature of SharedKD is its ability to dynamically select crucial structures based on gradients during the training process, maintaining high accuracy while optimizing the model for efficient lightweight deployment. This research marks a significant step beyond existing methods that reduce the size of already trained AI models post-training, allowing for the simultaneous development and training of both large and lightweight models.

By designing the models to aid each other's learning, SharedKD significantly reduces the computational load and operation time required for 3D perception models in high-accuracy and real-time scenarios, such as autonomous vehicles. This technology is expected to be widely adopted in future vehicle AI and edge artificial intelligence systems.

The research aligns with Korea's Ministry of Education's 'KMU Vision 2035: EDGE' strategy, particularly in the convergence of 'AI+X' and 'Mobility' areas. By applying AI model optimization techniques to the autonomous driving sector and collaborating with industries, this research showcases the university's targeted approach to shaping its research focus.

The DAC, a globally recognized international conference for electronic circuits and design automation, features a diverse range of research from semiconductor design to AI/ML systems and efficient AI implementation.