Minnesota DOT: traffic signs inventory
Case overview
RoadAR was provided with:
Disclaimer: The current work is a prototype made with the main goal to understand the future requirements and to showcase what RoadAR can deliver in potential future projects with MnDOT.
- MnDOT traffic sign assets GIS database
- Multiple drives of a limited geographical scope (2.5 miles, one way)
Disclaimer: The current work is a prototype made with the main goal to understand the future requirements and to showcase what RoadAR can deliver in potential future projects with MnDOT.
In 2021, Roadly delivered a pilot project with NYS DOT. We have built a full inventory of road reference markers for Albany County.
Challenges
NYSDOT has collected high-resolution photo imagery with the GPS coordinates for the roads within Albany County over several years. The content was collected by survey vehicles making shots every 26 feet (8 meters). Therefore it was challenging to detect the road markers with precise location.- Another challenge was to identify all reference markers, correctly recognize the text and other attributes and accurately locate each asset
on the map. - Not all the text was clearly visible on the provided imagery, so in some cases it was necessary to analyze the sequence of markers to determine the text value.
- Also it was crucial to discover which markers have been damaged over the years.
The solution
While delivering the project, we have built the algorithms for marker detection, text recognition, and localization.
We used a combination of neural networks to detect
reference markers. After the object detection neural
network finds a candidate it’s re-checked by a
classification neural network to minimize the number of false positives. On the next stage we detected 4 points that set the corners of the panel.
We used a combination of neural networks to detect
reference markers. After the object detection neural
network finds a candidate it’s re-checked by a
classification neural network to minimize the number of false positives. On the next stage we detected 4 points that set the corners of the panel.
Later on the text recognition algorithm was looking for characters inside the panel. Lastly, our algorithms
analyzed sequences of recognized reference markers
and highlighted possible mistakes for manual verification.
To understand themarker location, we automatically
estimated the camera position for every image frame
where a marker is present using our proprietary
algorithms based on the SLAM approach.
Finally all the detects across different shots of the same asset and different drives of the mapping vehicle were aggregated in one coherent map layer. That allowed us to increase the quality of the final output even more. We performed a thorough QC process for every step. In addition to that, we have done an independent verification of all the quality attributes of the produced asset inventory in order to ensure that it meets our quality standard and
the client’s requirements.
analyzed sequences of recognized reference markers
and highlighted possible mistakes for manual verification.
To understand themarker location, we automatically
estimated the camera position for every image frame
where a marker is present using our proprietary
algorithms based on the SLAM approach.
Finally all the detects across different shots of the same asset and different drives of the mapping vehicle were aggregated in one coherent map layer. That allowed us to increase the quality of the final output even more. We performed a thorough QC process for every step. In addition to that, we have done an independent verification of all the quality attributes of the produced asset inventory in order to ensure that it meets our quality standard and
the client’s requirements.
The outcome
1000 MILES - 2000 REFERENCE MARKERS FOUND
We identified and located over 2,000 reference markers while processing more than 1,000 miles (over 1,600 km) of the road network.
We identified and located over 2,000 reference markers while processing more than 1,000 miles (over 1,600 km) of the road network.
x4
MORE PRECISELY
We have localized
reference markers
based on estimation of
camera position for every
image. The average absolute
positional accuracy
of the markers was as high as
6.25 feet (1.9 meters).
- 99%
TEXT RECOGNITION ACCURACY
After implementing our AI
optical character recognition
and automatic markers’
sequence analysis followed
by manual validation for most
complex cases, we achieved
99% text recognition
accuracy.
over 96%
MARKER DETECTION RATE
We have achieved over 96%
marker detection
rate.