Transfer Learning-based Vehicle Classification

In this paper, we propose a transfer learning-based vehicle classification from the convolutional neural network (CNN) pre-trained on a large scale dataset. It is possible to construct deep neural networks effectively for new problems with a limited scale vehicle dataset. The proposed system is divided into two stages. First, the vehicle area is detected on the roadway video by Haar-like features. Second, the transfer learning-based vehicle classification using GoogLeNet classifies vehicle models. Experimental results show that the proposed system has a high accuracy of 0.983, which is 0.326 higher than that of the conventional method without transfer learning.

I. Introduction

Vehicle classification is a key element of the intelligent transportation system and has various applications such as traffic flow statistics, intelligent parking systems, and driver assistance systems [1]. In the past, several researches have been done on the vision-based vehicle classification using support vector machine (SVM) [2] to train classification models. The conventional method is not robust due to unstable feature extraction from illumination change. Because convolutional neural network (CNN) learns itself without the need for people to extract features, it complements drawbacks of the conventional method and has contributed to the rapid development of the image classification [3].