Traffic sign detection and recognition using deep learning-based approach with haze removal for autonomous vehicle navigation

• Published in: e-Prime Vol. 7; p. 100442
• Main Authors: Rani, A. Radha, Anusha, Y., Cherishama, S.K., Laxmi, S. Vijaya
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2024; Elsevier
• Subjects: Convolutional neural network; Deep learning; Haze removal; Traffic sign detection and recognition; Deep learning; Traffic sign detection and recognition; Convolutional neural network; Haze removal
• ISSN: 2772-6711; 2772-6711
• DOI: 10.1016/j.prime.2024.100442

The major contributions to this work are as follows:•Implementation of HRU-Net model for removing the haze from input images, and also pre-processes the image.•Implementation of TSDR-CNN model for detecting the location of traffic sign from haze removed images. Autonomous vehicle navigation technology is increasing rapidly. However, automatic sign recognition in complex illumination environments like low-light, hazy regions is a significant challenge in in-vehicle navigation. So, haze removal and robust traffic sign detection and recognition (TSDR) are critical for ensuring the vehicle's and its passengers' safety. However, the conventional methods failed to perform both haze removal and TSDR operations simultaneously. Further, the conventional haze removal methods eliminate the wanted pixels, still the presence of haze, which results in reduced traffic sign detection performance. Moreover, the conventional sign recognition methods classify a few types of traffic signs. So, this work aims to develop a unified model for multi-class sign recognition in complex environmental conditions. Therefore, this work introduced the deep learning model for haze removal based on TSDR (DLHR-TSDR). Initially, the CURE-TSD dataset is considered. The haze removal U-network (HRU-Net) module inputs a hazy image and outputs a haze-free image trained to learn the mapping between hazy and haze-free images. Then, the TSDR-convolutional neural network (CNN) module takes the haze-free image from the previous module as input and outputs the location traffic signs in the image. The simulation results on the Carleton University Retinal Eye-Traffic Sign Dataset (CURE-TSD) dataset show that the DLHR-TSDR method developed in the study resulted in 99.01 % accuracy, higher than traditional methods.