Jeong Jin-woo, a student in Professor Kim Jangho’s research team at the Department of Artificial Intelligence at Kookmin University, developed “SR-BEVS,” a camera-based 3D object detection technology resilient to environmental changes, through collaborative research with student Ahn Sang-ho and researchers from Hyundai Motor Company. The study was presented at IROS 2026, a world-renowned international conference on robotics.
 

Camera-based artificial intelligence for autonomous driving can experience degraded recognition performance due to factors such as fog, changes in lighting, and sensor noise. To address this, the research team proposed a method that involves removing certain connections within the model in different ways, performing restoration learning, and then averaging the weights of multiple models to integrate them into a single model. This approach improved detection performance without requiring additional inference operations, while maintaining the structure and size of the existing model.

Experiments using the nuScenes autonomous driving dataset confirmed performance improvements under various image degradation conditions, such as fog and sensor noise. Notably, under severe Gaussian noise conditions, the comprehensive detection performance metric (NDS) improved by 33.8% compared to the existing Fast-BEV model, and detection performance was also enhanced in typical images.
 
This research is significant in that it simultaneously considers both the ability to adapt to environmental changes and the computational constraints of real-world systems. It is expected to be utilized as a visual perception technology for autonomous vehicles and mobile robots in the future.
​
Meanwhile, this research was conducted through an industry-academia collaboration between Kookmin University and Hyundai Motor Company, and the paper is titled “Sparse Recovery Bird’s-Eye View Model Soup for Robust Camera-Based 3D Object Detection.”

(Photo: Student Jeong Jin-woo)