AI detection of unaccounted buildings for taxation purposes
This post is devoted to computer vision technology to identify buildings in urban areas, developed by our team for the State foundation.
The purpose of the technology development is a highly accurate and automated searching for illegal or unaccounted buildings for their subsequent involvement in the tax turnover.
In fact, this technology is a part of an integrated project «Property tax collections improvement» (https://lnkd.in/d-ZdDxW5 ), aimed at increasing the rate of the land tax collection.
From technical point of view, we solved the problem of increasing the accuracy of object recognition in comparison with neural networks processing aerial imageries of the territory. For this purpose, we used dense clouds of terrain points obtained as a result of photogrammetric processing of aerial photographs in the LAS file format as data sets.
Eventually, we worked out computer vision technologies based on training neural networks of the Mask RCNN and Unet architectures. Eventually, we developed an ultimate solution which performs the entire set of necessary operations (reading LAS files, converting point clouds into the input data format for neural networks, converting the response of neural networks into geometric objects in the real world coordinate system).
The Mask RCNN architecture was chosen as the final architecture due to its greater efficiency. The final accuracy of recognition the contours is more than 0.9.
The developed software was completely built in the “Property tax collections improvement” system (https://lnkd.in/d-ZdDxW5 )
Shapes of the buildings are used for automatic comparison with land cadastre and taxation authority data to identify buildings and sites that are underpaid in tax.
A significant part of them was subsequently involved in the tax turnover and it brings an income to the budget on a regular basis.
This post is devoted to computer vision technology to identify buildings in urban areas, developed by our team for the State foundation.
The purpose of the technology development is a highly accurate and automated searching for illegal or unaccounted buildings for their subsequent involvement in the tax turnover.
In fact, this technology is a part of an integrated project «Property tax collections improvement» (https://lnkd.in/d-ZdDxW5 ), aimed at increasing the rate of the land tax collection.
From technical point of view, we solved the problem of increasing the accuracy of object recognition in comparison with neural networks processing aerial imageries of the territory. For this purpose, we used dense clouds of terrain points obtained as a result of photogrammetric processing of aerial photographs in the LAS file format as data sets.
Eventually, we worked out computer vision technologies based on training neural networks of the Mask RCNN and Unet architectures. Eventually, we developed an ultimate solution which performs the entire set of necessary operations (reading LAS files, converting point clouds into the input data format for neural networks, converting the response of neural networks into geometric objects in the real world coordinate system).
The Mask RCNN architecture was chosen as the final architecture due to its greater efficiency. The final accuracy of recognition the contours is more than 0.9.
The developed software was completely built in the “Property tax collections improvement” system (https://lnkd.in/d-ZdDxW5 )
Shapes of the buildings are used for automatic comparison with land cadastre and taxation authority data to identify buildings and sites that are underpaid in tax.
A significant part of them was subsequently involved in the tax turnover and it brings an income to the budget on a regular basis.