Article Preview
TopIntroduction
In recent decades, vehicle classification has played a vital role in intelligent transportation systems, because the usage of vehicles has become increasingly universal in human life due to the rapid development of society. For vehicle detection and classification, existing methodologies utilize various types of information such as radar signal and acoustic signature (Zhou & Cheung 2016). The performance of these methodologies is vulnerable to several environmental variations such as weather, illumination noise and so on (Jiang et al. 2017).
With the proliferation of cameras, an abundance of video data related to vehicles, such as traffic on highways, road intersections, toll booths, etc. has become available for analysis, insights and even real-time actions. For example, in traffic measurement and management, vehicle detection and classification delivers important information and also assists in road planning and maintenance by understanding the distribution of dissimilar vehicle classes (Javadi et al. 2018; Wang et al. 2017). Also, vehicle detection and classification have become a research area in video-based intelligent transportation systems (Mithun et al. 2012). For example, counting the vehicles in busy intersections helps in reducing the level of traffic congestion. Due to all the above-mentioned reasons, there has been a significant amount of research undertaken, which addresses the challenging task of vehicle classification with the help of image/video data related to vehicles.
The research work on vehicles has received significant interest among researchers includes the applications such as fine-grained vehicle classification, vehicle detection, vehicle identification, vehicle classification and so on (Ke and Zhang 2020; Rachmadi et al. 2018) from image/video data. The fine-grained vehicle detection and classification becomes a challenging problem (Yang and Lei 2014; Li et al. 2019) due to the low interclass and high intra-class variance of images. Many classifiers have been used for vehicle classification such as AdaBoost algorithm (CHEN et al. 2018; Wen et al. 2014), dynamic Bayesian network (Kafai and Bhanu 2011), support vector machine (Şentaş et al. 2018), artificial neural fuzzy inference system (Murugan and Vijaykumar 2018), invariant Charlier moments (Aqel et al. 2017), etc. The vehicles have unique visual and structural characteristics compared to other objects, and also have small inter-class distances and larger intraclass variation. These factors make the detection and classification of vehicles difficult (Yu et al. 2017; Biglari et al. 2017).
In recent decades, Deep Neural Networks (DNNs) have been used extensively in image classification problems. The DNN classifier is the best choice for vehicle type classification when the additional prior information about the images is unavailable. Although DNNs have the advantage that the images can directly be fed as inputs, i.e. DNNs can play the role of feature extractor and classifier combined, this requires a large amount of training data and involves significant training time and computational resources. The requirement of training data size is reduced using strategies like transfer learning, where the deep neural network-based solutions still involve significant computation. While it is hard to come up with a single compact set of powerful descriptive features for complex image classification tasks that global-level descriptive features such as Histogram of Oriented Gradients (HOG) and Gabor filters have been used successfully in object detection tasks. By using these aforementioned feature descriptors instead of the raw images, the amount of training data can be reduced and the performance vehicle image classification can also be improved.