Capturing Speed Data through AI and Video Analytics
22 March 21
The monitoring of vehicle speeds (as well as volumes) is standard practice for data collection companies, these are an integral part to most traffic modelling software and are needed for transport planning purposes for new altered junctions. Whilst there any currently many methodologies for capturing this data, not many are perfect and have a number of limitations.
In partnership with Vivacity Labs, Tracsis have developed an innovative video-analytical solution for capturing both volume and speed utilising AI technology which is set to revolutionise the market. This method is suitable for post processing by running video footage through the system, and addresses some of the main limitations to standard data capture. These features are outlined in this article by Nick Mather of Tracsis.
What are the current methods available?
The most common piece of equipment for the capture of vehicle flows and associated speed is the pneumatic tube counter. Tube counters provide a very cost effective and reliable way of collecting this data.
Radar classifiers provide a similar output to pneumatic tube, albeit in vehicle length bins rather than individual classifications, and of course speeds. Again the usual limitations apply and set up proficiency is key to capturing good data using this methodology.
Inductive loops and AI sensors provide a permanent solution for longer term monitoring with “live data”, which can be viewed through a client portal such as an online dashboard.
ANPR can provide speeds over longer distances in the form of journey times, and other technology such as Bluetooth can provide similar data but passively as a sample of the total traffic.
Current challenges of temporary surveys
Looking at the main methodologies of pneumatic tube and/or radar classifier, and overlooking the standard limitations as mentioned earlier in this article, there are some additional major challenges that apply to this equipment.
High Speed Roads
The installation of tube on higher speed roads is not possible in the vast majority of cases. The width of the road and potential number of lanes is the minor issue, with the major problem being the higher speeds impact on the tube and the risk of the tube breaking, and the safety risks this poses. Flailing tube is also usually responsible for damage to vehicles and carries a high risk and a cost through insurance.
Physical presence of equipment
In the instance of a site on a high speed road, in most circumstances radar is utilised, where there is no physical presence of equipment in the carriageway, but rather attached to suitable street furniture. This removes the risk both for the physical installation itself but also due to being road-side.
Limited classifications for reporting
However, the switch to the use of radar also means reporting in length bins, rather than classifications, which is less than ideal for most models and for most onward reporting requirements.
AI Speeds- how does it work?
By utilising the same AI technology seen in a permanently fixed AI sensor, Tracsis are able to offer a separate, proprietary solution where video footage of any particular site can be post-processed through the system and accurately capture volume and speed of all vehicle classes as well as pedestrians and cyclists.
The AI software detects, classifies and tracks road users seen in the video. Two speed lines are drawn across the road environment to define the area of speed measurement. The software uses the real-world distance between the two lines and the width of the road to calculate a perspective transformation to map an object’s detected locations into real-world coordinate frame. From this, the speed of each individual object can be obtained.
Why use AI?
This method is safer than installing tube and can also be used where tube installation is not possible i.e. high speed roads. There is no requirement to enter the live carriageway so a camera can be installed from a road side position. Where the positioning of a radar device needs to be road side for the device to function correctly, a camera provides greater flexibility for installation points and it is possible that over bridges or similar can be used where the cost of traffic management outweighs the cost-benefit ratio of a project.
This method offers extended classifications over tube counters, e.g. cycles and pedestrians and means that where a radar would otherwise be used, length bins are a thing of the past and full classifications can be captured to align with the modelling software being used.
We can post-process existing footage from previous projects and provide the speed data
Active travel modes can be specifically captured, as well as their speed of travel
More insights are possible using the AI Portal, including path traces and social distancing data
How accurate is it?
We have carried out a range of tests on sites in a mix of environments, including residential roads, trunk roads and rural roads, tested against a pneumatic tube counter
In summary, data correlates with traditional methods and is within a 5% threshold, some of our findings are published below.
Case 1 – Free flowing 30mph – North of England
Here we took a 24hr period with a roadside mounted camera and a tube counter located at the exact geolocation. The test was carried out on 16th January 2021. The speed limit of the road was 30mph and conditions were dry and cold.
Over a 24hr period the average speeds of the ATC and AI Portal show a strong correlation and demonstrate the same profile. For example, at 18:00 both the ATC and the AI Portal show a slight decrease in average speeds, likely due to an increased traffic flow due to commuters. The speed summary table below also demonstrates this where both the ATC and AI portal produce comparable average speeds 28.27 +/- 6.37 mph and 28.90 +/- 7.54 mph, respectively.
The above graph shows the comparison between the speeds recorded by the ATC and the speeds recorded by the AI Portal. The data show a strong correlation (0.72) between the ATC and AI Portal speeds. There is a clear clustering of data around the line with an equal split of data above and below the line, indicating that the spread is due to the uncontrollable random error.
Case 2 – Busy residential road, 20mph – Central London
Here we applied the same methodology with equipment at the exact location, for this trial we also compared a radar classifier, as well as the tube counter. The test was carried out on the 21st January 2021 and conditions were dry.
We can observe that whilst the profiles of speed match across the 24hr period, there is a marginal difference in average speeds, with the tube reporting around 20mph, radar 23mph and AI Portal 19mph across the full survey period.
Similar to Case -1, the comparison here shows a strong correlation (0.93) between the ATC and AI Portal. The split of the data above and below the line indicates that differences are due to uncontrollable random error.
The data presented in these case studies demonstrates that the AI Portal is as effective at producing volume and speed data as the industry standard pneumatic tube surveys. The differences between the two methods are explained by the uncontrollable random error.