| Title |
Energy-based Out-of-distribution Detection Network and Diffusion Model-based Dataset for Soiling Detection in Rear Camera Scenes of Autonomous Vehicles |
| Authors |
문승훈(Seunghun Moon) ; 전창렬(Chang-Ryeol Jeon) ; 이정훈(Junghoon Lee) ; 정동혁(Donghyuk Jeong) ; 염석주(Seokju Yeom) ; 강석주(Suk-Ju Kang) |
| DOI |
https://doi.org/10.5573/ieie.2023.60.11.95 |
| Keywords |
Deep learning; Energy-based model; Diffusion-based model; Autonomous driving; Classification |
| Abstract |
It is a crucial problem in autonomous driving scene to detect whether a rear-view camera of a vehicle is soiled. The false detection of soiled contaminants as road objects makes it hard to operate autonomous driving. However, due to the absence of a soiling detection-specific dataset, recent research to solve the issue is not actively conducted while they show insufficient performance for real-world applications. To address these limitations, we propose an energy-based out-of-distribution (OOD) detection network and diffusion model-based task-specific dataset. In the experimental results, the proposed network reaches 94.73% of soiling detection accuracy, which is a 15.52% improvement compared to the baseline. In addition, training procedure with the proposed task-specific dataset further impoved the soiling detection accuracy to 95.24%. Finally, we embedded the proposed network onto a mobile device and conducted experiments to validate its real-time capability in autonomous driving scenarios. These experiments analyzed the trade-offs between the backbone-specific soiling detection performance as well as inference time. |