Detection of LPG fitted vehicles for risky areas with computer vision techniques

Vehicle use is becoming more widespread day by day due to the world population growth. Within the scope of intelligent transportation systems, the IT sector and the transportation sector work in an integrated manner to solve the problems caused by the increasing number of vehicles. Data obtained with sensors and cameras are analyzed with artificial intelligence-based information technologies and used in autonomous vehicles, security, traffic management, navigation and passenger information systems. Computer vision techniques enable machines to extract meaningful patterns and relationships from images by combining image processing and deep learning technologies. This technique, which is based on imitating the human brain, becomes capable of making a decision as a result of the patterns and relationships it learns. Computer vision techniques are widely used in many fields such as tourism, agriculture, health, industry, defense, transportation, service, e-commerce, etc. The applications developed in these areas provide solutions to various challenges. One of the challenges in transportation is related to vehicles using Liquified Petroleum Gas (LPG) fuel. Due to the flammability of the gases inside LPG tanks and the possibility of explosion, it poses a danger especially in certain areas in cities. Entry of LPG vehicles is prohibited in institutions and organizations such as hospitals, shopping malls, hotels, military campuses that have indoor parking services. The control method of the ban is carried out by assigning a personnel and checking the vehicle luggage by the personnel in charge. The aforementioned inspection method is not a sustainable system due to the necessity of employing the personnel in appropriate conditions, the margin of error of the personnel and the loss of time for the vehicle drivers. In this thesis, computer vision techniques are used to automatically detect vehicles running on LPG fuel. Vehicle image data captured by mobile cameras in different cities in Turkey are trained and compared with four different deep learning models. As a result of the training and performance tests performed on the models, the YOLOv8 model was more effective than the other models with an accuracy of 0.994 mAP and a speed of 11.6 ms. It has been shown to be a stable model in terms of real-time tracking and accuracy in real life. It is envisaged that the developed system can contribute to the applications of computer vision techniques as well as benefit the national economy, public life safety and environmental protection.