A non-intrusive category identification method based on the binary image of profile vehicles and CNN classification algorithm.

Automatic vehicle classification is one of the most critical subsystems in the Intelligent Transport Systems, as it is responsible for the validation of vehicle categories in Electronic Toll Collection Systems. The accuracy of the classification in this kind of process needs to be high. Therefore, it is necessary to use several intrusive sensors in traffic lanes and sophisticated classification algorithms in practical applications. This article proposes a new classification method based on binary images of vehicle profiles extracted from a non-intrusive optical barrier sensor. We use an AlexNet convolutional neural network as a vehicle classification algorithm. The last layers of the network have been modified for the vehicle classification domain. In the training process, we use transfer learning and data augmentation techniques. The proposed method was tested using data collected at a toll plaza in the Sao Paulo State Highway, Brazil. The first experiment was carried out, with 11,233 images grouped in 11 categories, resulting in a classification accuracy of 98.02%. In the second experiment, we used 194,361 images collected over four days. We were able to evaluate the performance of the method in adverse conditions, specifically heavy rains. In this scenario, the proposed method reached an accuracy of 96.41%, which indicates that the replacement of a set of intrusive sensors with only an optical barrier can be a viable alternative for vehicle classification. Additionally, we become available all the images data used in this study in Mendeley's open repository.