Robust Vehicle Classification Based on the Combination of Deep Features and Handcrafted Features

Vehicle classification plays an important part in Intelligent Transport System. Recently, deep learning has showed outstanding performance in image classification. However, numerous parameters of the deep network need to be optimized which is time-consuming. PCANet is a light-weight deep learning network that is easy to train. In this paper, a new robust vehicle classification method is proposed, in which the deep features of PCANet, handcrafted features of HOG (Histogram of Oriented Gradient) and HU moments are extracted to describe the content property of vehicles. In addition, the spatial location information is introduced to HU moments to improve its distinguishing ability. The combined features are input to SVM (Support Vector Machine) to train the classification model. The vehicles are classified into six categories, i.e. large bus, car, motorcycle, minibus, truck and van. We construct a VehicleDataset including 13700 vehicle images extracted from real surveillance videos to carry out the experiments. The average classification accuracy can achieve 98.34%, which is 4.49% higher than that obtained from the conventional methods based on "Feature + Classifier" and is also slightly higher than that from GoogLeNet (98.26%). The proposed method doesn't need GPU and has much greater convenience than GoogLeNet. The experimental results have demonstrated that for a specific task, the combination of the deep features obtained from light-weight deep learning network and the handcrafted features can achieve comparable or even higher performance compared to the deeper neural network.