Appropriate CNN architecture and optimizer for vehicle type classification system on the toll road

Abstract

The growth in the number of vehicles in Indonesia causes traffic jam problems, including on the toll roads. Traffic jam problem on the toll roads occurs because the users must stop and make payments in the toll gate. Government built an Automatic Toll Gate Shelter (or Gardu Tol Otomatis/GTO) as an effort to reduce this problem. However, GTO can only be used by certain type of vehicles only. In this study, we developed a system that can classify type of vehicles so that GTO can be used for various types of vehicles that cross the toll road. The developed system will receive vehicle input to be classified. The learning process to do the classification is using the Convolutional Neural Network (CNN). The CNN algorithm is trained first with 2,930 vehicle images divided into 1,794 vehicles type 1 (van, jeep, and pick-up) image, 507 vehicle type 2 images (truck with 2 axle), and 631 vehicle type 3 images (truck with 3 axle). From the experimental results of CNN architecture and various parameters of the architecture, the best accuracy is found on MiniVGGNet architecture which applies Adadelta optimization function and input image parameter 64x64 and epoch 40. The result obtained from the network has accurate evaluation or out-sample accuracy of 73%.