Intelligent vehicle recognition system using deep learning / Soon Foo Chong

• Main Author: Soon , Foo Chong
• Format: Thesis
• Published: 2019
• Subjects: TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://studentsrepo.um.edu.my/10309/1/Soon_Foo_Chong.pdf; http://studentsrepo.um.edu.my/10309/2/Soon_Foo_Chong_%E2%80%93_Thesis.pdf; http://studentsrepo.um.edu.my/10309/

Vehicle recognition is a challenging task and has a great global demand in various areas. In this research work, the vehicle recognition focuses on two main objectives, i.e. the recognition of the vehicle’s manufacturer and also its model. Generally, an intelligent vehicle recognition system which is mainly based on deep learning method, i.e. deep Convolutional Neural Network (CNN), has been presented to tackle both the recognition of vehicle manufacturer and its model. Targeting on the recognition of vehicle manufacturer, vehicle manufacturer logo images have been utilized as the key feature to be processed in the CNN. To recognize vehicle manufacturer logo images, two different approaches have been introduced to improve the recognition performance of CNN. Firstly, Zero Component Analysis (ZCA) whitening transformation technique has been adopted as the pre-processing step in the CNN system. ZCA whitening transformation technique is implemented to remove redundancy of adjacent image pixels. Experimental results show that after implementing ZCA-CNN, the vehicle logo classification accuracy is further improved from 99.07% to 99.13% over 10 vehicle manufacturers in the Xia Men University (XMU) dataset. Secondly, Particle Swarm Optimization (PSO) is utilized to optimize the hyper-parameters searching process for CNN architecture. Based on multiple PSO iterations, a set of best CNN hyper-parameters is selected to achieve the optimum vehicle logo classification result. The experimental results explicitly prove that the proposed PSO-CNN approach outperforms most of the state-of-the-art methods, achieving an accuracy of 99.1% over 13 vehicle manufacturers in the XMU-PLUS dataset. It is noteworthy that the vehicle model recognition task becomes more challenging due to the massive amount of vehicle manufacturers and their large intra-model variations around the world. Most of the existing vehicle model recognition methods focus on locating a global feature or extracting more than one local subordinate-level feature from a vehicle image. Hence, the Principal Component Analysis Network- based Convolutional Neural Network (PCNN) has been proposed. Only one discriminative local feature of a vehicle, which is the vehicle headlamp, is pinpointed for vehicle model recognition. The proposed model eliminates the need for locating and segmenting the headlamp precisely. In particular, PCNN leverages the effectiveness of both principal component analysis and CNN in extracting hierarchical features from a vehicle headlamp image and also reducing the computational complexity of the traditional CNN system. To further enhance the training procedure while still keeping the discriminative property of the network, the fully-connected layer is updated by using backpropagation optimized with Stochastic Gradient Descent (SGD). The proposed PCNN method is validated using PLUS dataset that comprises 13,300 training images and 2,660 testing images, respectively. The model is robust against various translational and rotational distortions. Experiments show that PCNN outperforms state-of-the-art techniques with an average accuracy of 99.51% over 38 vehicle makes and models using the Malaysia North-South Highway (PLUS) dataset. In addition, the effectiveness of the proposed method is also validated using the public CompCars data set, achieving 89.83% accuracy over 357 vehicle models.