الفهرس 
Chapter 1 IntroductionOnly 14 pages are availabe for public view
 |  | from | 131 |  |  |
| | | from | 131 | | |
Abstract
Smart cities and intelligent transportation systems (ITS) are the attention of many researchers right now. One of the key components of ITS that is crucial to many parts of daily life is vehicle recognition. This inspiration served as the basis for the thesis, which gives a thorough investigation into vehicle recognition, particularly fine-grained car recognition using deep learning. It compares four state-of-the-art convolutional neural networks (CNNs) which are Inceptionresnetv2, Inceptionv3, Resnet50, and Mobilenetv2 then builds an analysis based on a thorough examination of their architectural designs. The thesis also elaborates on some modifications to the state-of-the-art CNNs in order to enhance the accuracy by finding and paying more attention to the discriminating features between the car classes. The thesis then proceeds to introduce different training techniques that can be used to train the networks. TensorFlow was used to implement the CNNs, which were trained using the Stanford Cars Dataset. The accuracy acquired by the provided approaches produced satisfactory results; the Top 1 accuracy attained was 0.8851, 0.8819, 0.8523, and 0.8284 using our modified Inceptionresnetv2, Inceptionv3, Resnet50, and Mobilenetv2 respectively. We also demonstrated a pipeline that can be used to identify vehicles in videos. You Only Look Once (Yolov3) was used for object detection, the centroid tracking algorithm was used to track the vehicles, and we were able to achieve an average Intersection Over Union (IOU) of 0.7, which is thought to be sufficient to predict the general location of the vehicles in the videos.