Details

Type

Bachelor Thesis
Student Project
Master Thesis

Persons

1-3

Motivation

If we have an existing 3D scan/model of a scene, we want to point our smartphone there and see any changes as an augmented reality overlay immediately.
For this, we detect which objects have been removed/added or simply moved around. In many environments, it is useful to observe such changes, e.g. for inventarization of offices, warehouses, or urban spaces. Furthermore, this information can be used to train deep learning algorithms for the semantic understanding of such environments.

Description

Since 3D scans can be done in real-time using Kinect or smartphones, we also want immediate results of the change detection, unlike previous work which may take seconds or even minutes. To achieve this, we will exploit a hierarchical data structure for the scanned data so that the change detection can be run in parallel. In order to differentiate occlusions from real changes we need to know which volumetric regions of the space have already been seen by the sensor. For that, we employ a smart flattened data structure since its interior is empty. If we consider the known sensor device noise properties, we can also get robust results. Therefore, we will augment the data with the uncertainty from the 3D scanner so that we know how far the samples are allowed to diverge.
Here is a paper detecting changes in point clouds and giving an overview of related work: https://perso.telecom-paristech.fr/boubek/papers/ChangeDetect3D/ChangeD…

Tasks (extent depends on PR/BA/DA)

  • Smartphone app prototype that registers a point cloud (from RGB using depth estimation with machine learning, e.g., LiteDepth https://arxiv.org/abs/2209.00961) to a previously scanned point cloud
  • Based on a modified Infinitam (infinitam.org) implementation which uses Kinect 3D scanner to compare an existing scan in real-time with a live scan.
  • Store view frusta of sensor (the seen space) in a flattened data structure for differentiating seen changes from the unknown: https://www.cg.tuwien.ac.at/research/publications/2014/Radwan-2014-CDR/
  • Extend the comparison between octree nodes between the old and new model, based on the permitted distance (points + uncertainty ellipsoids), to classify into changed/unchanged geometry
  • Evaluate the robustness of the algorithm by using a Blensor virtual scan of a ground truth 3D model
  • Classify changed objects with machine learning
  • Track whether added and removed objects are the same object
  • Fast parallel implementation in CUDA and compare run-time to the state-of-the-art
  • Estimate volume change (using surface reconstruction and voxel difference)
  • Visualize changes over a time series

Requirements

C++ programming skills and interest in geometry processing. Experience in geometry processing and 3D data structures such as octrees, point clouds, machine learning, or CUDA will speed up the development tasks.

Environment

Platform-independent C++.

A bonus of €500/1,000 if completed to satisfaction within an agreed time-frame of 6/12 months (PR/BA or DA).

Responsible

For more information please contact Stefan Ohrhallinger, Michael Wimmer.