Modeling of Convolutional Neural Networks for Detection and Classification of Three Vehicle Classes

Abstract

The increasing number of vehicles causes traffic density on each road segment, especially in urban areas. Planning and regulating vehicle traffic, requires traffic flow parameter data, which currently is still largely done manually by assigning several people to count vehicles passing divided by a certain time period. The development of artificial intelligence in the field of computer vision enables the observation of traffic data using smart cameras. This study aims to produce a Convolutional Neural Netwoork (CNN) model for the detection and classification of three vehicle classes, namely bus, car and motorcycle. The architecture used to build CNN uses smallVGGNet with 3x3 convolution layers and one fully connected layer. The first three convolution layers are considered as feature extraction layers while the last one is for classification where there are three output types of vehicles. The model is trained using 3, 000 different images according to the class. In the training phase, it was determined 75 epochs where the consumption time of each epoch was ± 30 minutes so that the total time needed was 37.5 hours. The results of tests that have been done show that CNN models that have been trained can classify vehicle types with an accuracy of between 63-100%.