CNN-based Traffic Sign Detection on Embedded Devices
[ 1 ] Wydział Inżynierii Mechanicznej, Politechnika Poznańska | [ SzD ] doktorant ze Szkoły Doktorskiej
[2.2] Automatyka, elektronika, elektrotechnika i technologie kosmiczne
2022
rozdział w monografii naukowej / referat
angielski
- Traffic sign detection
- Mapillary Traffic Sign Dataset
- YOLOv4
EN Traffic sign detection is a key task in autonomous driving. In addition to high accuracy, the algorithm must operate in real-time on an embedded device. Traffic signs are often found occupying a small area of a high-resolution image and can be easily confused with other signs and billboards. We analyze the aforementioned challenges, using the YOLOv4 model, which we train on the Mapillary Traffic Sign Dataset (MTSD) with a designed data augmentation method and weighted loss function. We achieve AP50 = 59.0% on the validation dataset. The contribution of this work is a quantized YOLOv4 traffic sign detector with an input resolution of 960 × 960px. The proposed model is optimized to achieve better performance on devices with limited computational resources. Our model runs at 11.2 FPS on Nvidia Jetson Xavier AGX.
108 - 111
CC BY (uznanie autorstwa)
witryna wydawcy
ostateczna wersja opublikowana
w momencie opublikowania
20