Araç konum tahmini ve araç sınıflandırması derin evrimsel sinir ağları kullanarak

The aim of this master's thesis is to classify the vehicles and estimate the position with license plate localization using Deep Convolutional Neural Network (DCNN). Vehicle pose estimation with license plate localization serves as one of the most widely-used real-world applications in fields like toll control, traffic scene analysis, and suspected vehicle tracking. Along with license plate information, to obtain overall comprehension, the information of the owner vehicle also plays a great role, and contextual information is defined as the relationship between the vehicles pose license plate and the owner vehicle in our work. We proposed a one-stage anchor-free vehicle classifier for simultaneously localizing the region of license plates and vehicles' poses. The classifier, rather than bounding rectangles, gives bounding quadrilaterals, which gives a more precise indication for vehicle pose estimation with license plates localization. For single scale input, we reached mean Precision Accuracy mAP/mAP50 of 35.4/82.3 on the Laboratory for Intelligence and Safe Automobile (LISA) benchmark dataset, already outperformed the existing commercial systems OpenALPR and Sighthound. For multi-scale input, we reached the best mAP/mAP50 of 40.8/90.1. For the vehicle pose (front-rear), classification accuracy reached 98.8%, average IoU reached 71.3%, giving a promising result as an end-to-end vehicle position estimation and license plate localization with contextual information. The work has performed in python programming language with several libraries of deep learning were being used for this purpose. There are three functional head networks in our design, and thus the end-to-end and simultaneous training leads to a potential training instability. If the model fails to converge, it is quite hard to trace which head design is raising a problem. Thus, in our design process, we added each the functional head step-by-step, making sure the single head design idea is working correctly and then expanded the model with the rest functional heads. Our DCNN model training started from an initial weight which we had already trained for about 110000 iterations in the model without classification head, so the total training iterations will be around 780000 including the transfer learning part in DCNN. Transfer learning made the DCNN model start at a smart point and made it easier to optimize all of the functional heads simultaneously. Keywords: Vehicle classification, pose estimation, optimization, DCNN, transfer learning, license plate, localization, deep learning.