Supervision: Nicolas Aspert

Project type: Master thesis


Airborne Mapping provide nowadays a huge amount of data from different sources such as imagery or LiDAR point cloud.The Classification of LiDAR point is conventionally done using geometric routine (triangulation, neighboring...) but since few years, "AI" routines appeared to extract classes of point clouds (ground, building vegetation, power lines...) Nevertheless, the percentage of correct classification rarely reach 80% and thus the remaining 20% requires either te use of geometric routine and manual correction (time consuming) As point cloud does not contain significant thematic information (except intensity of return), the AI routines will depend mostly on the point cloud density that has been used for training. And the AI service/application available on the market generally trained their model on open source LiDAR data (OpenDEM, Swiss Alti3D, LiDAR HD IGN...)

Those open dataset are generally national coverage with low/medium resolution and thus using those routine for higher resolution samples is generally problematic. In combination to the LIDAR data, the imagery acquired simultaneously is also a source that is currently not exploited in the "AI" approach for classifying point clouds. As the position and orientation of the image is known accurately, the correspondence of LiDAR point with a pixel in the image is easy to obtain.

The purpose of the project would focus on multimodal learning to improve point cloud classification. We propose to focus on Power line dataset that contains also oblique images (forward/backward) and high density point cloud (>60 pt/m2)

This project will be done in collaboration with the Helimap company. Sixense Helimap is a company active in airborne mapping. We use sensor integrating LiDAR sensor with different sets of high resolution camera (> 100Mpix). We are focused on corridor mapping (linear infrastuctures) such as power line, railways, highways, pipeline.