Traffic sign recognition system for Indian road conditions using YOLOv4

Abstract

Traffic Sign Detection and Recognition is a vital topic for autonomous driving vehicles, cruise control and adaptive driver assistance systems. Real-time prediction of the traffic signs captured from car-mounted cameras and dashcams is a challeng ing task. Because the images captured may contain scenarios like low illuminated images, the small size of signs relative to the image, motion blur, object occlusion, and degraded/discolored traffic signs. Also, the perspective of traffic signs can be different. To address the issues, we introduce India Traffic Sign Dataset (ITSD) to improve the performance of models on Indian Roads. We also propose improve ments in the existing YOLOv4 model. First, we use the Generalised Intersection over Union (GIOU) distance metric for Anchor box calculation. Second, we introduce an extra feature scale layer for detection. Third, we change connections from the CSPDarknet-53 backbone to the PANet neck to improve the utilization of shallow features. Finally, we modify the final feature extraction step (Detection Block) and use Grouped Convolutions to represent features better. We also employ an image illumination model to tackle the problem of low-light images captured at night. The proposed model is tested on three existing datasets, namely Mapillary Traffic Sign Dataset (MTSD), Tsinghua-Tencent 100k (TT100K), and German traffic sign de tection benchmark (GTSDB). We achieve a 94.73% mAP on the TT100K dataset, which outperforms the existing state-of-the-art models. We conduct ablation on each of our proposed modifications, and finally, we perform cross-dataset validation to test the robustness of our system. Keywords: Traffic Sign Detection and Recognition; YOLOv4; Object Detection; Computer Vision; Dataset