Automatic Vehicle type Classification with Convolutional Neural Networks
Max Naegeler Roecker, Yandre M. G. Costa, João Luiz Ramalheira de Almeida, Gustavo H.G. Matsushita · 2018
This paper proposes a convolutional neural network model for classification of vehicle types with low-resolution images from a frontal perspective. This characteristic can be useful to the development of systems with limited resources, like embedded systems. We trained the model as a multinomial logistic regression where cross-entropy of the ground truth labels and the model's prediction estimates the error. To prevent overfitting, we performed data augmentation in the training dataset and regularized the model using the dropout method. Experimental results in a subset of the BIT-Vehicle Dataset with samples uniformly distributed between classes shows that the model achieves an accuracy of 93.90%. We conclude that the model is discriminative and capable of generalizing the patterns of the vehicle type classification task.