PROJECT TITLE :

DFR-TSD A Deep Learning Based Framework for Robust Traffic Sign Detection Under Challenging Weather Conditions

ABSTRACT:

The development of a reliable traffic sign detection and recognition system, also known as TSDR, is of the utmost significance for the productive application of autonomous vehicle technology. Because of how important this task is, a significant amount of research has been done on it, and as a result, the available literature contains a number of promising method proposals. On the other hand, the majority of these methods have been evaluated on clean and challenge-free datasets, and the performance degradation that is associated with various challenging conditions (CCs) that obscure the traffic-sign images captured in the wild has been ignored by the majority of these evaluations. In this paper, we investigate the TSDR problem while considering CCs, with a particular emphasis on the performance degradation caused by these considerations. In order to accomplish this, we suggest a prior enhancement focused TSDR framework that is based on a convolutional neural network (CNN). Our modular approach is comprised of a CNN-based challenge classifier, Enhance-Net, which is an encoder-decoder CNN architecture for image enhancement, and two separate CNN architectures for sign-detection and classification respectively. We propose a new training pipeline for Enhance-Net that focuses on the enhancement of the traffic sign regions (rather than the whole image) in the challenging images subject to their accurate detection. This pipeline focuses on the enhancement of the challenging images subject to their accurate detection. In order to determine how effective our method is, we used the CURE-TSD dataset, which is made up of traffic videos that were recorded under a variety of CCs. We demonstrate through experimentation that our approach achieves an overall precision of 91.1% and a recall of 70.71%. This represents an improvement of 7.58% and 35.90% in precision and recall, respectively, when compared to the existing benchmark. In addition to this, we evaluate our strategy against a variety of CNN-based TSDR methods and demonstrate that our strategy is superior to the others by a significant margin.