AI-based Traffic Management

Is AI-based traffic management on its way to becoming standard in Denmark? Effective traffic signal control using AI and object-based traffic data can help reduce congestion, improve traffic safety and decrease emissions while contributing to meeting the UN's climate goals by promoting sustainable urban mobility. The prospects are promising.

Bent Seerup

Illustration of the principle behind SWARCO Technology's SMART AI

Management of traffic signal systems in Denmark has become increasingly advanced. Today, almost exclusively "smart" sensors are used, which are capable of generating large amounts of data that can be used directly to control the signal change in modern traffic signal systems. Artificial intelligence (AI) to process the large amount of traffic data has in recent years been applied in the development of signal control algorithms that can more or less "predict" the need for green time and optimise waiting times within the system's signal groups.


Traditionally, communication between detectors and the control unit of a signal system has occurred by sending simple detector outputs (on/off) to the control unit's I/O inputs. After that, the control unit processes the various inputs based on the logic and criteria established in the traffic signal programme.

With modern sensors that have builtin data processing, the criteria for the inputs sent to the traffic controller can be made conditional on vehicles' stop line distances, speeds, expected arrival time (ETA) and classification (size). By allowing the traffic controller to process the sensors' 'raw' data instead of letting this happen in each individual sensor, control of the intersection can be based on the total amount of data available. Each sensor constitutes an independent data source, where the location, movedetermined based on the sensor's position (local coordinate system). Real-time data from each individual sensor must be collected, validated, formatted and assembled in a structured form.

Faulty data must be removed and replaced with valid ones, 'duplicates' reduced, etc. This process is, of course, very complicated and requires in-depth interface knowledge. In addition to the detector data the traffic controller receives through its associated sensor system, data can also be collected from car GPS systems. Many car manufacturers send data from their vehicles to cloud-based systems.

This is part of the development of modern car technology, which allows manufacturers to collect and analyse data to improve the car's performance, safety, and user experience. In the long term, access to these data could be included in the data within the system's signal groups. ment, etc. of each individual object is foundation of signal control.


The reinforcement learning (RL) model used for optimising the green time allocation in a traffic signal system can essentially be described as follows:

1. The model's optimisation goal is to allocate green time to the signal group (or groups) that has the greatest "impact" on reducing the total weighted delay in the system.

2. The model continuously calculates the current total weighted delay based on recorded and expected data. This is done by summing the  delays of individual objects distributed across signal groups.

3. The model's sole action is essentially to decide which signal group (or groups) should currently be green to achieve the optimisation goal. A signal group can be suspended until a threshold for maximum waiting time for that signal group is reached, after which the signal group should be activated.

4. The model monitors the development of the system's average delay and "learns" to recognise certain patterns, which it uses in its decisions.


is a branch of artificial intelligence where an algorithm learns to make decisions by receiving feedback as a result of its decisions/actions. The algorithm receives rewards or punishments based on its actions and seeks to maximise its accumulated reward over time.


are two different technologies used for detecting and tracking objects in traffic sensors.

LIDAR uses laser light to measure distances by emitting short pulses of laser light and measuring the time it takes for the light to be reflected back to the sensor. This time is measured very precisely, providing accurate distance measurements.

RADAR uses radio waves (electromagnetic waves) to measure distances and speeds by sending out radio waves and measuring the delay and Doppler effect of the returned waves. This makes RADAR more suitable for measuring speeds. 


include targeted areas of effort to reduce traffic-related environmental nuisances and promote sustainable mobility. These goals include reducing air pollution, noise pollution, and CO2 emissions to improve the urban environment and public health. The UN's goals also include initiatives to reduce traffic accidents and create safer traffic conditions through traffic safety measures and education.


cost society a lot of money each year in terms of welfare losses, lost life years, hospital expenses, home care and rehabilitation, sick leaves, etc.

  • A traffic fatality costs about 5 million €.
  • A seriously injured person costs about 1 million €.
  • A slightly injured person costs about 100,000 €. Source: Transport Economic Unit Prices, DTU 2022


The algorithm/model allocates green time 7 The model 'learns itself' which data to the various signal groups based on an optimisation goal to reduce the total system's average waiting time. By assigning different weights to certain directions and/or traffic groups, prioritisation can be made, as the optimisation will then be based on the system's average weighted waiting times. Thus, for example, prioritisation of 'soft' road users and/or public transport can easily be made. The model can register the actual computational passage time for each signal group as input for the optimisation algorithm.

belong to each signal group, meaning the model itself determines lane distribution and stop lines. Therefore, the model adapts to roadwork where traffic lanes are reduced or moved. If a sensor's direction is changed or replaced, the model selfcalibrates/corrects the sensor's input, etc. SWARCO Technology is currently testing their SMART AI algorithm in two existing facilities in Denmark, and preliminary results suggest that even in newer traffic-controlled facilities, improvements in average waiting times of more than 5-10% can be achieved by letting the AI optimisation algorithm decide the distribution of green time.


In addition to optimising the distribution of green time, traffic signal control must also be able to end green phases safely. Through the analysis of object data, traffic controllers should extend the green light as long as there are two or more vehicles in the same lane within the selection zone. This means within a distance from the stop line and at a speed where it is possible to both stop and continue through an amber light. This approach can reduce the risk of rear-end collisions. If the traffic signal system is located in a rural area, where the speed limit is 70 km/h or higher, the traffic control should also be able to extend the green light as long as there is a vehicle more than four seconds away from the stop line but closer to the stop line than its current stopping distance (the dilemma zone). This can reduce the risk of red-light running collisions (lateral collisions). The signal control should also be able to delay the activation of conflicting signal groups if the intersection area is not cleared (inter-green extension). Examples of this type of 'event detection' could in8 clude left-turning traffic, pedestrians on the crosswalk or on the roadway, etc. If LIDAR sensors or thermal cameras with object tracking technology are used for detection in the intersection area, the algorithm can detect near-miss incidents based on the relative positions, speeds, and directions of the objects.


With the use of AI-based traffic management, there is no need to load a signal programme into the traffic controller. Once the algorithm is developed, it can in principle be used generally. SWARCO's SMART AI uses a parameter-based user interface, where the basic prerequisites and control parameters of the system are entered. This includes information about:

  • The system's signal groups
  • The interdependencies/dependencies of the signal groups- Conflict/inter-green time matrix
  • Minimum green times for the individual signal groups
  • Maximum waiting times for the individual signal groups
  • Prioritisation (weighting) of the individual signal groups and/or traffic classes

If the control device is connected to the internet or integrated into a monitoring system, for example, prioritisation assignment to the individual signal groups and effectiveness monitoring can be performed via an online dashboard. The parameter-based user interface provides easy access to modification and adaptation of the control algorithm, thereby also ensuring a high degree of supplier independence.

SMART AI Dashboard with real-time and historical data: Current traffic flow and overview of cycle times


Typically, an existing traffic signal system can be upgraded with sensors featuring object tracking technology and AI-based control for a cost between 20,000 and 40,000 €, depending on whether the existing control unit can be reused. An investment of this size can often be recouped within a few months, solely in terms of expected savings from accidents. Additionally, there are long-term effects on climate and public health. Since the control of the traffic signal system is self-optimising and self-monitoring, the need for 'human' intervention is limited to physical maintenance. Monitoring and optimising traffic signal systems using artificial intelligence (AI) has significant advantages that are not possible to achieve manually. Especially in times when human resources are often scarce, AI can play a crucial role in improving traffic signal systems in several ways.

The use of AI enables automatic realtime monitoring and rapid response to changing traffic conditions. While humans can monitor traffic, it is often difficult to respond instantly to sudden traffic jams, accidents, or roadworks. AI systems can analyse large amounts of data from sensors, cameras, and other sources in seconds and adjust the traffic controller in real time. This results in faster reactions, less waiting time, and reduced delays for road users.

Using AI takes into account complex traffic patterns and variables that humans might overlook. AI algorithms can analyse historical data, weather conditions, realtime events, and even data from vehicles and their GPS systems. This allows for signal control adjustments specific to time periods and locations, maximising traffic f low. Additionally, AI can prioritise traffic safety in ways that humans may find challenging. AI systems are capable of detecting dangerous situations and potential conf licts between vehicles and pedestrians. This allows for extending green time at critical locations, reducing the risk of accidents, and improving overall traffic safety.

Finally, AI can operate around the clock without needing breaks or rest. This means that traffic signal systems are constantly monitored and optimised, resulting in more efficient signal control and reduced waiting times, regardless of the time of day or year. This level of constant optimisation is simply not realistic to achieve manually.

In the end, the use of artificial intelligence leads to better performance of traffic signal systems, increased convenience for road users, reduced waiting times, and improved traffic safety. It is an investment in both efficiency and convenience that helps to tackle the challenges associated with increasing traffic loads in modern cities. In 2024 alone, SWARCO Technology expects to have commissioned 10-20 AIbased traffic controllers in Denmark and approximately 15 in Norway.