Computer vision technology with machine learning (ML) is becoming increasingly popular for collecting road inventory data and identifying road defects cost-effectively. As the ML model cannot identify all the defect types required for maintenance planning, it is supplemented by the Rapid Condition Assessment tool, which allows manual click for capturing location, images, recorded voice, etc., allowing inspection to be undertaken without stopping the vehicle. With an integrated treatment selection and optimisation tool, an optimal routine maintenance works program can now be prepared/updated quickly.
When this technology was used for recurrent road safety inspections to satisfy the Victoria Road Management Act 2004, we identified an issue with the accuracy of GPS coordinates affecting defect location between subsequent inspections. This issue created some difficulties in filtering out defects captured earlier. To ensure that only new defects were added to the database, an advanced data analytics process was developed and used to correct the GPS coordinate issues. A user interface was added to the application to quickly review images and playback videos to check the defect type and duplicates before creating a new defect record.
Another hurdle that we successfully solved to make the system operationally viable was the large volume of media data management. Data uploading (large media files in the low-speed internet environment at regional depots) and processing of data captured by inspectors from numerous contracts were optimised and streamlined for consistent and timely output.
Our experience shows that the new inspection approach of object recognition technology can significantly reduce inspection time, improve the accuracy of data collected, and help reduce missed defects/hazards. A time series of pavement distress data also helped identify rapidly deteriorating road sections and prioritise the sites for timely maintenance with the right treatment.