Artificial intelligence helps build roads


Artificial intelligence helps build roads

For the first time in Russia, transport and economic surveys were carried out using artificial intelligence. Within the framework of the design of the “Northern Bypass of the city of Perm” facility, carried out by JSC “Institute Gipromstroymost”, the traffic intensity is recorded by means of video recording from an unmanned aerial vehicle and subsequent analysis of video materials using a software package.


The analysis of “modern” methods of accounting for the intensity of movement was carried out in sufficient detail by Professor M.R. Yakiv in his monograph “Transport planning: creating transport models of cities” 2013. On the example of research carried out in Perm, Professor Yakimov examined three main methods used in practice.

The methods were evaluated according to the following indicators:
one-time costs;
current expenses;
time spent on collecting / processing information;
quality assessment of collected data.

The analysis showed that the third method is the most expensive from the point of view, since it requires the installation of sensors under each lane. Accordingly, the first method is the least expensive. The processing of the materials collected by the first method took 1 hour, the second – 3 hours, the third – 15 minutes. The second method is the most accurate, as it allows to count cars from video recording several times in order to eliminate the factor of human error. Therefore, the semi-automatic method was taken as a reference. When compared with the standard, the first method showed relative deviations from 3 to 33.5%, the third – from 7 to 57%. Such errors of the third method are caused by inaccuracies in reading information and imperfection of the devices available to researchers.

Conclusion: comparative analysis showed that to obtain initial information for single intersections, it is necessary to use a semi-automatic method of collecting information is impossible at the intersection under study, then it is advisable to use a full-scale data collection method. Data collection using transport accounting sensors showed that the data received was unreliable, the errors amount to more than 20%.


With the advent of quadrocopters in our life, the method of counting with video recording has become even more relevant, since now we are not tied to the places where the traffic surveillance cameras are installed, and we can cover the entire intersection, which allows us to simultaneously record the movement of cars in all directions. However, the processing of such video materials presents a certain difficulty. For example, it turned out to be rather difficult to determine the traffic intensity in each of the 9 directions at the roundabout in the alignment of General Butkov Street (Kaliningrad).

Considering that in order to determine the uneven distribution of the traffic flow per unit of time, it was necessary to analyze the video filmed at different times of the day for several days, then one can imagine what labor costs this entailed.


The solution to the problem was modern technologies of computer vision and machine learning. The TrafficData software built on the basis of the latest convolutional network architectures made it possible to bring the semi-automatic method described by M.R. Yakimov to a fully automated one. Once again, artificial intelligence has made it possible to save a person from routine tedious work.

As mentioned above, the new method was first applied during the survey of traffic flows on the Northern Bypass of Perm, carried out in October 2019 by the TrafficData development team on the instructions of the Giprostroymost Institute.

In the course of this work, the survey of the road situation was carried out by quadrocopters at eight registration points simultaneously. 200 flights were carried out, 50 hours of video were processed, thanks to which it was possible to obtain the intensity during peak hours and deduce the real distribution of intensity during the day. Thanks to the TrafficData software, which allows you to analyze such a volume of video information in the shortest possible time, there is now no need to use the non-uniformity coefficients to determine the average daily intensity given in ODM 218.2.020-2012 “Guidelines for assessing the traffic capacity of highways.” It is known that these coefficients are determined on the basis of measurements made on the Moscow Ring Road with the help of sensors (see the thesis for the title of Ph.D. “Regularities of changes in the intensity of urban car traffic”, Mendeleev G.A., 2001) … Essentially, this data is obtained locally and cannot be used universally.


Let’s take a closer look at how the task of analyzing traffic flows in TrafficData software is solved. After the video is ready, there are only 4 simple steps left to get the result:

  1. Downloading the application – Install on the working computer
  2. Uploading video – Start car recognition
  3. We indicate the gates – To count cars in different directions, we process the gates
  4. Uploading the results – By the intensity and composition of traffic flows by directions


  • Accuracy of vehicle recognition over 90%
  • Definition of 18 types of vehicles according to SP 34.13330.2012 and SP 396.1325800.2018
  • Processing video from ground cameras
  • UAV video processing
  • Processing video from night vision cameras
  • Availability of entrance, exit and through sections with the direction of accounting of cars in one and in both directions
  • Determining the direction of vehicle travel (trajectory tracking)
  • Tracking a vehicle passing under a bridge and local obstacles (trees, poles, billboards) without breaking tracks
  • Adaptive gates – gates follow limited image movements
  • Determining the speed of cars
  • Processing speed: for RiPNB 2 times faster than video playback speed
  • Availability of tools for improving the quality of video recognition: tool for connecting tracks, tool for fixing the type of car
  • Output of results in Excel
  • Possibility to select coefficients of reduction to a passenger car according to SP 34.13330.2012, SP 396.1325800.2018, as well as custom
  • Calculation of daily intensity according to ODM 218.2.020-2012
  • Generation of video with processing results

It is very easy to check the accuracy of the calculations – just watch the video generated by TrafficData with the results, which clearly displays the count of cars when driving through the gate, types of cars and even their speed, and make sure that no car is overlooked


TrafficData is used as a data collection tool for designing highways, junctions, bridges, city planning, and obtaining initial data for macro- and micro-modeling of traffic flows.

In order to implement the state policy in the field of traffic management, on April 18, 2019, the order of the Ministry of Transport of the Russian Federation No. 114 “On Approval of the Procedure for Monitoring Road Traffic” was issued. In accordance with the order, it is necessary to collect data on traffic conditions for the entire road network annually! It will cost the country enormous funds if it does not learn how to automate this process. TrafficData is the closest one on the way to this goal today. For this, the following functions are planned to be implemented next year:

  1. Determination of the transport port delay at the intersection;
  2. Determining the length of the queue of cars
  3. Accounting for pedestrians.

With the new technology for collecting traffic data faster, more economical and more accurate, and with the upcoming workload of monitoring the road network, there is confidence that the use of intelligent software like TrafficData will gain wider acceptance. And after a while, no one will remember what they used to do differently.