On the Development of Indian TSR Systems using Machine Learning Techniques

Abstract

Advanced driver-assistance systems (ADAS) is a typically growing aspect of the automobile industry. Famous automobile brands like Brenzo have introduced various automated configurations like Traffic sign recognition (TSR), cruise control, speed control, crewless vehicles etc. TSR refers to systems or techniques that aim at automatic recognition of traffic signs. Traffic signs are markers at the side of roads that instruct or inform the drivers about the road rules and conditions ahead of them. TSR is targeted at reducing human-car interaction and aimed at automated systems in vehicles for the recognition of traffic signs. It is needless to say that the current death rate in the globe due to traffic accidents is alarming. The same situation prevails in India as well. Therefore, automated systems like TSR has become indispensable part of ADAS. This thesis aims at developing efficient practically implementable architectures emphasizing on datasets. In this line, an Indian dataset has been proposed to strengthen the research further. In this context, four different schemes have been proposed for TSR architectures. In isolation, each scheme has been validated using three standard databases, namely GTSRB, BTSC and IRSDBv1.0. The achieved accuracy is promising classification accuracy compared to the existing schemes elaborated in each chapter. In this thesis, the second chapter presents a wholesome review of the related works. To eliminate the lack of an Indian traffic sign dataset, the first contribution presents a fully annotated Indian traffic sign dataset named IRSDBv1.0 which is now available in public domain. The second contribution aims at designing an efficient Deep neural network (DNN) scheme, namely MDEffNet. MDEffNet achieves high accuracy on GTSRB and BTSC despite having low number of parameters. But MDEffNet didn’t give expected results in case of IRSDBv1.0 dataset due to limited image samples probably. The following chapters are in line with these findings. The third contribution explores the efficiency of wavelet descriptors. Along with the descriptors, three classifiers, namely CNN, CNN ensemble and LSTM, have been deployed to analyse the effectiveness of the proposed scheme. The fourth contribution eliminates the bottleneck created by wavelet descriptors by using curvelet descriptors along with Shannon entropy in its framework. Besides increasing the attained accuracy, this approach reduces the feature vector length considerably. All the presented architecture works on the assumption that the input images are free of visual challenges like occlusion, blur etc. Considering occlusion being the most common visual challenge, in fifth contribution a modified ResNet50 DNN architecture has been proposed, namely GDNN, that performs well on partially occluded signs. All the three datasets viz; GTSRB, BTSC, and IRSDBv1.0 have been used to validate the proposed schemes. Finally, the overall concluding remark is laid out in the last chapter along with future directions.