Improvement of Traffic Measurement AI by Utilizing Digital Twin

Generally, AI training data for object detection purposes consists of footage captured in real-world settings. In this study, we propose using footage generated from a digital twin constructed within a game engine as training data, instead of using actual footage. We evaluated the performance using night-time footage, where the pretrained model, trained on the COCO Dataset, exhibited significantly poor accuracy. By training the model with images created using the method developed in this study, we observed a substantial improvement in Average Precision (AP) from 69% to 88% compared to the pretrained model. Additionally, significant improvements were also observed at other test locations. Based on these results, it is considered that this method is effective in enhancing the performance of object detection.