Description

Objects like roads and buildings change over time. These changes can be detected in time series of images or point cloud data. In the Netherlands we can make use of a sequence of highly detailed laser scanner data, called AHN data with a point density of about 10 points per square meter. This data contains XYZ information, but also class labels, so it is known to which object class each point belongs to. Recently, deep CNNs have demonstrated their superior performance in extracting representative features for various computer vision tasks. As a specific CNN architecture, Siamese networks perform well in applications which compute similarity or to detect changes between two inputs. In this MSc research you will analyze the working of a Siamese CNN network to detect relevant changes of objects.

Objectives and Methodology

Literature review on change detection using multi epoch point clouds, Siamese networks. Implement suitable deep learning network, generate training data and change detection on multi epoch dense point clouds. Analyze object changes, finetune training data and network parameters, perform accuracy assessment.

Further reading

https://www.mdpi.com/2072-4292/11/20/2417/htm