Traffic sign-board detection in unconstrained environment using improved YOLOv4

Abstract

Traffic sign detection and recognition in an unconstrained environment is a chal lenging task for autonomous vehicle operations. Small traffic signs in the captured image make this problem harder. Further, a model needs to detect and recognize these signs accurately as this is a real-time problem. This work proposes a modified YOLOv4-based deep learning model that uses CSPDarknet53 as base architecture. 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 In tersection over Union (GIoU) as distance metric instead of Intersection over Union (IoU). In our modified architecture, we have proposed two different improvements in the YOLOv4 model. First improvement is an modified PANet with grouped con volutional layers in the detection neck, and an additional feature scale for detecting larger traffic signs is utilized. In the second improvement, we have used dense con nections in the detection block, which helpes in better accuracy with less inference time. The proposed model is experimented on Mapillary Traffic Sign Dataset (MTSD) and Tsinghua-Tencent 100K dataset (TT-100K). MTSD consists of global traffic signs from different countries, and TT-100K consists of traffic signs from China. Further, we have proposed our dataset, consisting of Indian traffic sign images, and tested the performance of these images. The proposed model is compared with existing state of-the-art models. We have achieved an accuracy of 94.80% on the TT-100K dataset, which outperforms existing methods. We have also performed the cross-data experi ment 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