Nanosemantic has developed detectors of road infrastructure facilities for Digital Roads

2024: Creating Software for Digital Twins

The developer of neural network solutions "Nanosemantics" has created software for digital twins of road transport infrastructure objects based on artificial intelligence for the company "Digital Roads." This was announced by Nanosemantics on June 10, 2024.

Digital Roads specializes in real-time spatial measurements and monitoring of urban and industrial infrastructure facilities.

Digital counterparts of the road infrastructure allow you to track the organization of traffic, as well as the state of traffic lights, signs, sidewalks and quickly make decisions to bring them into proper condition. In addition, such automation helps to more effectively use the time of specialists, which they previously spent on routine, and sometimes dangerous processes. To create digital counterparts of the infrastructure, artificial intelligence software ‒ detectors of road infrastructure objects are used, the development of which was carried out by Nanosemantics specialists within the framework of this project.

The project took almost two years and took place in ten stages according to the number of detectors presented, among which were such as MAP (small architectural forms), gates, barriers, road stands and others. In the process of creating a solution based on computer vision and machine learning technologies, the developers chose and implemented the best algorithm from those investigated in terms of accuracy and speed. For high-quality training of neural networks in each of the directions, training and validation datacets were formed, numbering several thousand images from six cameras of mobile laboratories. Geodata became a partner in marking the data provided for subsequent training of neural networks.

The most difficult was the creation of detectors where small objects were involved, and high recognition accuracy was required. For example, reflectors and cameras ‒ small objects, and often in the image they merge with the background. This task required careful preprocessing and preprocessing of images, high-quality preparation of the data set and the selection of optimal models for training the neural network.

In total, more than 77 thousand images were used for training, while the image processing speed was 3 bps per batch of 6 images (or 18 fps). According to the results of the final testing, all detectors showed compliance with the metrics established by the customer initially: box mAP - 0.55, mask mAP - 0.48, F1-measure - 0.83.