View Independent Vehicle Make and Model Recognition

Abstract

This thesis presents a broad study and comparison of classifiers for vehicle make and model recognition(VMMR) on a Pakistani Cars dataset. VMMR is a critical topic in the Intelligent Transportation System (ITS). In smart traffic and surveillance systems, sensors and cameras collect data. Before they can be used for real-time vehicle detection, videos acquired by these cameras are processed. For this, a variety of image-processing techniques are utilized. With these techniques, a surveillance system can overcome challenges such as arifacts, reflected images, weather changes, daylight variations, partial images , and change in perspective. The primary purpose of the proposed classifier is to overcome these challenges and improve the fine-grained vehicle classification results. In VMMR, multiclass classification is a significant challenge. There are two main categories in this situation: (a) plurality and (b) obscurity. The first mentioned problem is about different shapes of different vehicle models. We encounter the second issue in the case of when vehicle models are different and they are from different companies but their front/rear views are similar, or when car models belong to different manufacturers but they look the same. Most of the available large-scale image-based VMMR datasets are from developed countries. Hence, these datasets provide a good starting point for research purposes. However, traffic dynamics are very different and inconsistent for developing countries like Pakistan, especially on city thoroughfares. For instance, there are no lane markings more often than not, and driving within lanes is typically not the norm. Therefore, we are conducting our experiments on a Pakistani Cars dataset [101]. We conduct our research to check the validity of popular classification techniques on a Pakistani Car dataset and propose the best method. Our research involves deep neural networks as a feature-extracting tool. Pre-trained convolutional neural networks (CNN) are first fine-tuned on vehicle datasets. Features are extracted from the fully connected layer of the CNNs. Classification is done by (i) Fully connected layer (ii) FDLSR based classifier [102] (iii) SVM and (iv) KNN classifier. Experimental results show that the best classifier is Fisher Discriminant Least Square Regressiom(FDLSR) with Top-1 accuracy of 97.77% and Top-5 accuracy of 100% for features extracted from ResNet-152. Top-1 accuracy of 97.45% and Top-5 accuracy of 100% for features extracted from ResNet-50 and Top-1 accuracy of 90.91% and Top-5 accuracy of 100% for features extracted from VGG-16.