Analysis of Deep Convolutional Neural Network Models for the Fine-Grained Classification of Vehicles

Abstract

Intelligent transportation systems (ITS) is a broad area that encompasses vehicle identification, classification, monitoring, surveillance, prediction, management, reduction of traffic jams, license plate recognition, etc. Machine learning has practical and significant applications in ITS. Intelligent transportation systems rely heavily on vehicle classification for traffic management and monitoring. This research uses convolutional neural networks to classify cars at fine-grained classifications (make and model). Numerous obstacles must be overcome in order to complete the task, the greatest of which are intra- and inter-class similarities between the manufacturer and model of vehicles, different lighting effects, the shape and size of the vehicle, shadows, camera view angle, background, vehicle speed, colour occlusion and environmental conditions. This paper studies various machine learning algorithms used for the fine-grained classification of vehicles and presents a comparative analysis in terms of accuracy and the size of the implemented deep convolutional neural network (DCNN). Specifically, four DCNN models, mobilenet-v2, inception-v3, vgg-19 and resnet-50, are evaluated with three datasets, BMW-10, Stanford Cars and PAKCars. The evaluation results show that mobileNet-v2 is the smallest model as it is not computationally intensive due to depthwise separable convolution. However, resnet-50 and vgg-19 outperform inception-v3 and mobilenet-v2 in terms of accuracy due to their complex structure.