Road infrastructure is a critical asset for economic development. India now has approximately 62 lakh kilometres of road network. And timely maintenance has become another critical requirement.
Manual methods to detect and assess road damage are time consuming and require a person in the field. There are semi-automated methods which involve manual collection of field data and automatic assessment in the lab. Fully automated processes require sophisticated sensors like cameras and laser scanners for damage identification. All these methods are time consuming and expensive. Hence, most road networks have inefficient maintenance.
Recently, Deeksha Arya and team at IIT Roorkee collaborated with researchers from Japan to analyse the suitability of an existing Japanese automated model for identifying road damage. The method uses an algorithm to automatically detect road damage from images acquired by smartphones. Though the method seems to work in Japan, when it comes to roads in other countries, the performance is poor.
So the researchers collected nearly 3600 road images from the Czech Republic and 9900 from India using smartphones installed on vehicle windshields. Along with the images from Japan, they thus had a dataset of more than 26620 images where damages were annotated.
Different countries use different systems to classify road damage. The team chose four which were common: longitudinal cracks, transverse cracks, complex cracks and potholes.
Since the images varied in terms of resolution, the researchers rescaled them all to the same pixel resolution.
They used the dataset to train and evaluate 16 deep neural network models – data from each country separately as well as in combination, to check whether there were any significant differences due to the combinations. They also used transfer of learning from one network to another by creating ensembles of models.
As expected, models trained only with images from one country did not do well in recognising and identifying road damage in other countries.
The ensemble model, based on You Only Look Once version 5 (YOLO-v5) scored better than the other models.
“The models can automatically detect and classify road damage quickly using data from any country,” explains Durga Toshniwal, IIT Roorkee.
There are problems that remain, she admits. Road joints are often misclassified as cracks, and manhole or drain covers are identified as potholes. But these are challenges that can be overcome with a bigger dataset and more training.
“With a smartphone application for uploading geo-tagged images of damaged roads, drivers on the road could provide data on the current state of local roads. The automated system can then classify the damage and alert the concerned department to take appropriate action,” says Deeksha Acharya, IIT Roorkee, describing the future the team foresees.
Automation in Construction, 132: 103935 (2021);
National Remote Sensing Centre
STEAMindiaReports: Science news free for Indian media to use