Vehicle make and model recognition system for occlusion and bad lighting images

Abstract

Intelligent transportation system (ITS) is a massive and very significant sector in the socio-economic context of contemporary society. The need to use roads continues to increase, and this comes with the need to establish more efficient vehicle detection methods. Vehicle Make and Model Recognition (VMMR) has become an important aspect of vision-based systems, since it is applied to access control systems, traffic control, surveillance, and security systems, among others. However, the use of VMMR is challenging due to numerous factors, such as camera angle, poor lighting, and occlusion. Most of the existing works are focused on designing a VMMR system in a normal scenario, where the dataset is set for an ideal scenario, a scenario without illumination, or occlusion. Recent studies have used certain methods to extract the features by extracting the region of interest (ROI) of the front or the rear view of the vehicle to detect and recognize the vehicle. However, the aforementioned methods would fail with poor lighting or occlusion cases. In this thesis, a VMMR system is introduced, which begins by building the dataset, a combination of a benchmark dataset (dataset1) and a self-collected dataset (dataset2). A new approach of image enhancement method was applied to improve the low-light dataset. Then, the enhanced geographical feature extraction techniques were applied to extract the headlight and license plate. For occlusion cases, a new grid-based Speeded-Up Robust Features (SURF) was presented to extract the ROI even in the presence of an occluded object. Two classification approaches were used to recognize the make and model of the vehicle. The first approach is based on the Decision Tree, Support Vector Machine (SVM), and K-Nearest Neighbor (KNN) algorithms, where they are called ensemble classifiers, which can predict the VMM accurately. This is because the Decision Tree classifier predicts instances by sorting them based on feature values, while for SVM, the precision of classifying can be enhanced by employing suitable factors. As such, Radial basis function (RBF) kernel and optimized factors were chosen for SVM and KNN, where the testing data was classified by comparison to the k nearest training data based on a distance function. The second approach is the PCANet-II classifier, an approach with second-order pooling and binary feature variance with promising accuracy. The overall performance of the work in this thesis demonstrates a promising outcome, where the overall accuracy reached 96.08% by adopting an ensemble classifier and two datasets (dataset1, dataset2), while the PCANet-II classifier achieved 97.56% using both datasets (dataset1, dataset2). In conclusion, this approach proposed in this thesis showed higher performance than existing methods when bad lighting and occlusion are considered.