Artificial intelligence is going through a revival since the 2010s as the widespread of the Internet allow huge amounts of data to be exchanged. Deep learning algorithms that are a subset of machine learning algorithms use neural networks to predict results. These algorithms are trained on classified data samples to create a model. This model is then validated on classified datasets when comparing the predictions with the actual results. Once validated, the model is used on new datasets to obtain results for unclassified data. With time, the number of data samples used to create the model increase and therefore, the accuracy of the model improves, namely the quality of the predictions given by the model improves.

Mobile mapping systems capture large amounts of data from different sensors that could be used to train deep learning algorithms. In this article we will describe areas where AI could be integrated with mobile mapping technology.

Mission planning

Deep learning algorithms could be designed to predict the optimal time window for scanning. The inputs would be the size of the area to capture, weather forecasts predictions (for outdoor environments) and light exposure conditions. The model would provide the user with an indication of the best time to scan to ensure no rainy conditions and optimal light exposure for the cameras.

Trajectory planning

An AI model could create and then display in real-time the optimal trajectory to follow on the field. This would ensure exhaustivity of the data scanned, high density and shortest amount of time to capture a pre-defined zone. The inputs to feed the model could be a few pictures of the area to scan, a CAD plan, weather forecasts on the date of the mission, satellites configuration and some adjustable parameters such as density needed or specific points of attention to capture.

Example of trajectory planning
Example of trajectory recommendation

This step of mission planning could be done conjointly by field and office teams, each accounting for their own specific requirements, increasing and improving the communication between the two teams.

Identification of similar situation

To a human eye, every scanning configuration is new. But to a deep learning model taking into account geometrical features configuration, satellites alignment, distance from base stations, etc. similarities can be identified. Based on past experience, a performant algorithm trained on hundreds of previous field data captures could provide recommendations to the operator on how to operate adequately.

Field worker during mobile mapping campaign
Field worker during mobile mapping acquisition campaign

Real time feedback

Viametris systems already embed real-time feedback such as notifications about the number of satellites seen, live GNSS trajectory, point cloud display on the tablet, etc. While these inputs are already useful, added specific alerts triggered by detection algorithms could notify the operator about difficult areas to scan. They could also provide recommendations to facilitate the scanning. Sound or visual alerts would ensure that the operator adapts its scanning operation while ensuring safety. Examples: slowing down, adapting skeletal positioning, or changing side of the street for GNSS reception.

Classification on point clouds and images

Point clouds and images are data that can be classified. While images-based deep learning is well developed, point clouds-based models are less common. Geo-Plus provides powerful deep learning algorithms that can classify any kind of point clouds using intensity or RGB layer. This automatically classified point clouds can then be used in CAD, GIS or BIM applications to quicken the production of maps, drawing or BIM models.

Example of a point cloud classified using deep learning


Mobile mapping systems use various kinds of sensors, collecting vast amounts of data, which size can even increase when coupled with external parameters. The possibilities of integration with artificial intelligence are numerous and while intrinsically limited, artificial intelligence still demonstrates good performances at single tasking with precise goals. It could be a great leap forward in the years to come. Stay tuned!


Comments are closed