Deep Convolutional Neural Networks For Vehicle Classification

Vehicle recognition system plays an important part in Intelligent Transportation Systems and thus is a keen research topic as well. With the rapid development of image processing, pattern recognition and computer vision, more and more attention and research has been drawn to image-based vehicle recognition system. The image-based vehicle recognition system first captures the vehicle image from the video stream in camera, then pattern recognition and computer vision are employed to attain the valuable information in order to carry out classification and recognition. Nowadays, however, as constrained by speed and accuracy, intelligent vehicle recognition systems still witness few practical applications. Therefore, how to recognize the vehicle models efficiently and accurately is of great significance to guarantee the smoothness of the traffic.In this thesis, image-based object recognition is employed to implement the intelligent recognition of vehicle models. Firstly, as a public dataset is still unavailable up to now, a large number of vehicle pictures photographed on expressways are used to construct such a dataset. These images are abundant in complex backgrounds, changeable illumination and various scales, which ensure the diversity of the dataset. The final dataset of vehicles includes the training samples and test samples of vehicles such as cars, buses and trucks.Secondly, this thesis studied the classification and feature extraction methods. Combining with SVM algorithm, HOG and PCA-SIFT feature classifiers were trained based on the image data which was built in this paper and this thesis compared the performance of the classifiers.Finally, to improve the feature extraction speed and classification accuracy in existing image-based vehicle recognition systems, this thesis proposes a new method. This method applies Convolutional Neural Network (CNN) to extract feature independently, also combined with SVM to implement vehicle recognition. We optimize the network in terms of the layers of the network, sizes and the numbers of the filters and the selection of activation function and so on. Compared with the conventional HOG and PCA+SIFT features and algorithm for vehicle recognition studied in recent years, the approach in this thesis has considerable improvement in speed and accuracy.