Performance evaluation for vision-based vehicle classification using Convolutional Neural Network

Abstract

Vision-based vehicle classification is a very challenging task due to vehicle pose and angle variations, weather conditions, lighting qual ity, and limited number of available datasets for training. It can be applied for driver assistance system and autonomous veh icles. This paper conducted a performance evaluation for this task based on three Convolutional Neural Network (CNN) models, which are simple CNN, and pretrained CNN models that are AlexNet and GoogleNet. A dataset of more than 7000 images from the Image Processing Group (IPG) has been used for training and testing and the results indicate that AlexNet achieves the best classification result that is 65.09%. This result is obtained because of the variations of the quality of the images.