APPLICATION OF SIAMESE NETWORK AS A METHOD FOR VEHICLE COUNTING CLASSIFICATION BASED ON TRAFFIC VIDEO

• Main Author: Azali, Aldo
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75326
• Institution: Institut Teknologi Bandung

Vehicle counting is one of the Computer Vision technologies commonly used for traffic flow analysis in many developed countries. The application of Computer Vision in vehicle counting involves object detection, object classification, and object tracking methods. In each application, many vehicle counting algorithms have demonstrated good results and performance. However, until now, the applications of vehicle counting are limited to differentiating and classifying only the types of vehicles, and it has not been able to distinguish and classify the products and brands of the vehicles yet. Therefore, this final project proposes a vehicle classification method for vehicle counting using the similarity learning approach with objective able to classifying the product and brands of the vehicles. The selection of similarity learning as the vehicle classification method is based on the difficulty of obtaining datasets and the inefficiency time and resources of using traditional object classification algorithms for classifying vehicle products and brands, compared to using similarity learning. The similarity learning algorithm used in this final project is the Siamese Network. The Siamese Network algorithm is applied by calculating the similarity distance between an image of a branded vehicle A and another image of a similar branded vehicle A, as well as an image of a different branded vehicle B. The distance calculation is performed using the Euclidean distance formula. Additionally, the Contrastive Loss function is used as the evaluation metric for the Siamese Network learning. The obtained results show an accuracy of 98.45% on the training set (80% of the total data) and an accuracy of 90.79% on the testing set (20% of the total data) with a threshold distance < 0.5 as a parameter for considering two images as similar. Furthermore, the obtained results are used as the classification model for testing with a support set. However, the classification testing results show varying performance values depending on each tested vehicle product. The evaluation methods used for the testing results are the confusion matrix, precision, recall, accuracy, and F1-score.