| Title |
Improvement of Impaired Pedestrians Detection Performance using 3D Model-based Synthetic Image Generation |
| Authors |
이재용(Jaeyong Lee) ; 김학일(Hakil Kim) |
| DOI |
https://doi.org/10.5573/ieie.2023.60.2.56 |
| Keywords |
3D Model-based data; Synthetic image; Object detection; Impaired pedestrians; YOLO |
| Abstract |
When learning about objects in applications where securing existing learning data is very limited, problems such as data imbalance and lack of learning data can occur. This paper presents a method for generating real-life synthetic images using 3D models. The proposed technique is as follows. First, we render 3D models using a Blender, and obtain ground masking information using SegFormer, an image segmentation model after 2D image projection. Second, we generate labeling information and synthesize images similar to reality. The generated datasets were compared with the precision and recall when the YOLO-based models for real-time detection and the number of datasets were different, respectively. When 2,000 images were added sequentially to the experimental dataset and the contrast dataset, the detection of YOLOv4 in the proposed method compared to the contrast dataset increased by 16.73%p in mAP, 28.5%p, 26.19%p in Wheelchair and Person classes, respectively, but decreased by 4.3%p in Whitecane class. |