Application Of Visualization Of Convolutional Neural Networks In Advanced Driver Assistance System

In recent years,there are great improvement on GPU computing capacity,visual sensor technologies and computer vision.In the field of vehicle industry,smart transport system has become a hot issue.One of the most popular application is Advanced Driver Assistance System(ADAS).Meanwhile,deep learning techniques have brought impressive performance on various tasks,such as image classification,speech recognition and text analysis,etc.The commercial market pays lots of attention of the products using machine learning,deep learning techniques in order to implement into ADAS.Although deep learning is a very promising technique,the embedded systems has their constrains of storage and responding time limitation.That makes them hardly compatible with each other.In this paper,in order to solve this problem,we study the visualization state-of-the-art techniques of convolutional neural networks(CNN),which help to understand and diagnose the networks.We then develop a part of the visualization techniques and extent their contribution to optimize the architecture of CNNs.In this work,we propose the optimization tools and verify them in Traffic Light Recognition(TLR)dataset.In addition,we adapt the interpretation tools using visual analytics approaches to our dataset,in order to have the better understanding of the networks used.This report describes the study based on my experience of the corresponding project,which is organized as follows: the context of the project is presented in section 1.Section 2 summarizes the basic theory of CNN and the state-of-the-art research of visualization methods.Section 3 presents the extended development and application of the optimization tools using visualization.The adaptation of the model interpretable tools using visualization is described in section 4 and the technical overview is discussed in section 5.