Assessment of Vehicle Category Classification Method Based on Optical Curtains and Convolutional Neural Networks

Automatic Vehicle Category Identification (AVC) is essential for Electronic Toll Collection (ETC) Systems. Currently, various sensors and classification algorithms are used to achieve high accuracy in vehicle category classification. This work proposes a new classification methodology, based on non-intrusive sensors, which uses binary vehicle profile images. We used a convolutional neural network (CNN) based on AlexNet, which could be embedded in real-time computational systems. The CNN was trained using transfer learning and data augmentation techniques. The proposed methodology was tested using binary images of optical curtains installed in toll plazas, located on highways in the State of São Paulo, Brazil. Four experiments were carried out to evaluate the CNN classification accuracy under different operational and environmental conditions. In a controlled experiment, we used about 11,000 images, including original images and augmented images, to train the CNN with 11 vehicle categories. The method achieved an accuracy of 98.02%, very close to one of the current systems based on intrusive sensors. Additionally, we evaluated the performance of the methodology in another experiment, reproducing real operating conditions, involving about 194,000 images. The proposed methodology reached a global accuracy of 96.48%. The reduction in accuracy was due to classification failures resulting from interferences that occurred in the images, due to variations in environmental conditions, such as rain and night operation. The applicability of the proposed methodology can be extended to other contexts, such as the free-flow toll system, through the use of other types of sensors, such as LIDAR or cameras.