Traffic sign detection in unconstrained environment using improved YOLOv4

Traffic sign detection and recognition in an unconstrained environment is a challenging task for autonomous vehicle operations. The small traffic signs in the captured image make this problem harder. Furthermore, detecting and recognizing these signs accurately in real-time is crucial. This work proposes a modified YOLOv4-based deep learning model that uses CSPDarknet53 as the backbone. We have applied data preprocessing and image enhancement strategies for better model generalization. For this purpose, a nighttime image enhancement method is used to illuminate night images. In our work, prior to the YOLOv4 model, anchor boxes are calculated using the K-Means clustering algorithm, which uses Generalized Intersection over Union (GIoU) as the distance instead of Intersection over Union (IoU). Our modified architecture uses an improved PANet with grouped convolutional layers in the detection neck and an additional feature scale for detecting smaller traffic signs. The proposed model has been experimented on the Mapillary Traffic Sign Dataset (MTSD) and the Tsinghua-Tencent 100K dataset (TT-100K). MTSD consists of global traffic signs from different countries, and TT-100K consists of traffic signs from China. We have also tested the performance of the proposed model on our own dataset, consisting of Indian traffic sign images. The proposed model is compared with existing state-of-the-art models. We have achieved an accuracy of 94.80% and 80.71% on the TT-100K dataset and MTSD dataset, respectively, which outperforms existing methods. We have also performed the cross-data experiment on the German Traffic Sign Detection Benchmark (GTSDB) and Indian Traffic Signs Dataset (ITSD) using the model trained on MTSD. We have achieved 91.74% and 63.64% accuracy on GTSDB and ITSD datasets, respectively.