The modern society today boasts of the power of machine learning technology that ranges from web searches to recommendations on e-commerce websites. The machine learning systems help in the object identification of images, speech transcription and even in the selection of results of a search (Arel et al., 2010). The use of deep learning has made it possible to transfer the benefits of the machine learning technology to the transport system. The modern transport system demands the adaptation of the vehicles as well as its user interface to the road situation as well as the environment. Systems such as the lane departure warning system (LDWS) exist to make the work of drivers easy by alerting the drivers of any lane deviation (Abdić et al., 2016). Another example where the use of a machine learning system has worked to improve human lives includes maximising engine efficiency according to the road type.
Scholars have used different features to predict the driving environment. The features used include distance, standard deviation, average, maximum as well as the percentage of the time spent when speeding (Abdić et al., 2016, Taylor et al., 2012). Environmental perceptions have become a major research priority in driver assistance systems. Diverse sensors as well as fusion algorithms are the key aspects of autonomous driving. Therefore, machine learning approaches have increasingly found relevance in various environmental models in the time when the complexity of rule-based approaches has advanced. This is mainly true in urban environments where road classification has significantly gained popularity.

Order Now
Use code: HELLO100 at checkout

Road classifications ensure that drivers are aware of the type of roads to use. This knowledge provides several benefits, which include but are not limited to tailoring the vehicle interface to suit the particular road type. Also, autonomous driving functionalities only operate on roads that have structural separation of pedestrians and oncoming traffic. Deep learning methods consist of multiple layer representations that are obtained through the composition of non-linear modules which transform representations from one level to a higher level. Higher layers of representation are used for classification procedures. The layers allow for the amplification of input aspects that are critical for the suppression of irrelevant variations. For instance, an image is presented in the form of several pixel values. In the first layer, the learned features indicate the presence or absence of edges at various image orientations (Sikiric et al., 2014). The second layer enables the detection of motifs through the identification of particular edge arrangement. The third layer enables the subsequent layers to detect objects through assembling the motifs into larger combinations. These layers of features are suggested from the data collected for the general learning procedure and, as such, increase the credibility of the classification results.

Various research scientists have used several methods for road type categorisation. The road types have included freeways, local, arterial, and freeway ramps. For example, some authors have presented the use of artificial neural network, also known as ANN, to provide information to drivers on the type of road as well as the congestion level. Through the use of ANN, the authors, Tang and Breckon (2011), analyse the road through the use of Real-time images. Features such as texture, colour as well as the edge features are analysed from various sub-regions of the road (road edge, roadside, and road). Hybrid vehicles depend on the information about these aspects of energy optimisation. These authors, however, allowed for the comparison of two classification approaches by illustrating the performance of the noise-tolerant classification, ANN, and the discriminative nature of derived feature representation, K-Nearest Neighbours (k-NN). Unlike ANN, k-NN is a classification approach that is conducted based on categorising given instances.

From their analyses, Tang and Beckron (2011), suggest that the ANN approach has the potential of providing a more successful classification for two determination problems, namely the on-road and off-road determination problem. Despite the fact that k-NN increased the classification performance, it provided a classification of 70% which was lower compared to the ANN classifier of 90-97%. These results, however, contradicted the results reported by Açar and Bayir (2015), who indicated the completeness of K-Nearest Neighbours when compared to Naive Bayes classifiers. The use of ANN for the analyses is considered appropriate because of its high efficiency and success rate in various class problems such as on-road or off-road, multiple-lane motorway and off-road.

According to Sikiric et al. (2014), Tang and Breckon failed to consider the effects of poor lighting on the algorithm performance. Sikiric et al. (2014), chose to classify the traffic scenes by comparing various image descriptors. The authors assert that Spatial Fischer Vector (SFV) and Support Vector Machine (SVM) have the potential of providing the best performance in the presence of unrestricted bandwidth. However, a GIST image descriptor that is used with SVM classifier should be the best option when one is on a restricted budget.

Another type of classification is noted in the study conducted by Mioulet et al. (2013), where the authors used a Gabor filter bank to classify the type of the road. These authors extracted Gabor filters from various sub-regions of the image and analysed the Gabor filter responses through the use of histogram representations. Using the random forest approach, the scientists then conducted the road categorisation which achieved about 98% correct classification. The authors state that the Gabor filter approach is similar to the bag-of-word approach where low-level features are clustered to produce an understanding of an intermediate level.

Another group of authors proposed a novel content-based approach to the unsupervised learning of image features classification of road types. Slavkovikj et al. (2014), assert that the use of the novel content-based approach provides similar results to the state-of-the-art method that is used for classification of road surfaces. Taylor et al. (2012) used a data mining approach to classify the road type. Contrary to the approaches used by the other authors, Taylor and his colleagues chose to classify the road types using data mining of the vehicle signals. The authors compared the success rate of various classifiers for different road type classification. The road classifications included signage (None, White, Green and Blue), road type (A, B, C and Motorway), and carriageway type (Single or Double). From their report, the Random Forest algorithm indicated the best performance when used with the data summary of the wavelet Gaussian. Seeger et al. (2016) used an approach that classified the problem from the machine learning perspective and indicated that the use of pre-trained Convolutional Neural Network (CNN) in road type classification provided positive results when compared to training from scratch (Seeger et al., 2016).

Road type classification through deep learning promises great benefits for the transport industry. The benefits range from easy time on the road for the drivers, reduced time in traffic, to reduced road collisions. However, there is lack data on the types of classification approaches that are reliable to provide consistent results over time. Researchers in this field have reported varying results probably because of the different study objectives, different variables of data collected as well as different study conditions. From the contradicting results, the researchers have argued that there is no one approach that stands out above the rest. Therefore, different study objectives may require different approaches. Therefore, there is need for more research in this area to identify the best and most reliable method of classification.

    References
  • Abdić, I., Fridman, l., Brown, D. E., Angell, W., Reimer, B., Marchi, E. & Schuller, B. Detecting road Surface Wetness From Audio: A Deep Learning Approach. Pattern Recognition (ICPR), 2016 23rd International Conference on, 2016. IEEE, 3458-3463.
  • Açar, S. A. & Bayir, S. 2015. Road Detection Using Classification Algorithms. JCP, 10, 147-154.
  • Arel, I., Rose, D. C. & Karnowski, T. P. 2010. Deep machine learning-a new frontier in artificial intelligence research [research frontier]. IEEE computational intelligence magazine, 5,13-18.
  • Mioulet, l., Breckon, T. P., Mouton, A., Liang, H. & Morie, T. Gabor features for real-time road environment classification. Industrial Technology (ICIT), 2013 IEEE International Conference on, 2013. IEEE, 1117-1121.
  • Seeger, C., Müller, A., Schwarz, l. & Manz, M. Towards road type classification with occupancy grids. IEEE Intelligent Vehicles Symposium (IV) Workshop: DeepDriving-Learning Representations for Intelligent Vehicles, 2016. 1-4.
  • Sikiric, I., Brkic, K., Krapac, J. & Segvic, S. Image representations on a budget: Traffic scene classification in a restricted bandwidth scenario. Intelligent Vehicles Symposium Proceedings, 2014 IEEE, 2014. IEEE, 845-852.
  • Slavkovikj, V., Verstockt, S., De Neve, W., Van Hoecke, S. & Van De Walle, R. Image-based road type classification. Pattern Recognition (ICPR), 2014 22nd International Conference on, 2014. IEEE, 2359-2364.
  • Tang, I. & Breckon, T. P. 2011. Automatic road environment classification. IEEE Transactions on Intelligent Transportation Systems, 12, 476-484.
  • Taylor, P., Anand, S. S., Griffiths, N., Adamu-Fika, F., Dunoyer, A. & Popham, T. Road type classification through data mining. Proceedings of the 4th international conference on automotive user interfaces and interactive vehicular applications, 2012. ACM, 233-240.