This is a personal memo of this survey paper.

Introduction

Automatic change detection is an essential research subject.

Originated from Photogrammetry and Remote sensing.

  • Image-based methods deduced 3D changes from 1D or 2D measurements.

Point clouds are promising.

  • Can be obtained in a short time and a low cost
  • Free from perspective distortions
Four major domains cover ten categories of application topics.
Domains Categories
Urban monitoring land use and land cover (LULC)
building investigation
indoor variation analysis
vegetation surveys
Construction automation construction monitoring
infrastructure maintenance
historical heritage preservation
Hazard identification natural hazard monitoring
water body and flood monitoring
Cadaster cadaster updating
Three fundamental tasks must be addressed.
  • Coordinate system alignment
  • Spatial and spectral comparison
  • Change representation and analysis

Theoretical primer

Define and distinguish the term "change".
  • Binary definition: changed / no change
  • Triple definition: new / demolished / no change
  • Additions for partial change: partially new / partially demolished

However, definitions above are not sufficient to describe some complicated situations.

Define "change" semantically at object-level.

From the aspect of one object, these types follow the natural and intuitive way of human observation.

  • appeared / disappeared / fully moved
  • partially moved / deformed
  • unchanged
Change detection is still challenging.
  1. Inconsistent sampling
    • Why? Point clouds with different densities / distributions / covered regions.
    • Solvable with comprehensive observation
  2. Limited visibility
    • Why? Inevitable generation error due to occlusion / perspective.
  3. Missing semantics
    • Point cloud partition is feasible in some cases.
    • Semantics are required for object-level analysis.
The general objective of change detection is twofold.
  1. Spatial changes between two datasets at different timestamp from the same site.
  2. Change types derived from spatial changes.
\[\begin{gather} P,Q\subset\mathbb{R}^3\\ \Downarrow\\ M=\{(p_i,q_i,d_i) | p_i\in P^\ast, q_i\in Q^\ast, d_i\in D_{p,q}\}\\ \Downarrow\\ L=E(D_{p,q}) \end{gather}\]
  • $P, Q$: input point clouds
  • $P^\ast, Q^\ast$: overlapped subsets from $P, Q$
  • $D_{p,q}$: spatial difference
  • $E(\cdot)$: evaluation function
  • $L$: change types

General workflow

Change detection generally follows threes stages.
  1. Reference frame registration: Point clouds are aligned into the same coordinate system for comparison.
  2. Geometric difference estimation: Spatial differences obtained via occupancy analysis or distance measures.
  3. Spectral and attribute analysis: Changes identified by analyzing geometric differences and attribute shifts.
1st: Reference frame registration

Reference frame registration is an optimization problem for estimating transformation parameters for point clouds without markers.

flowchart TD PC1("Point cloud 1") PC2("Point cloud 2") STEP1("Correspondence matching") STEP2("Transformation calculation") OUT("Transformation parameters") PC1 & PC2--> STEP1 STEP1 --> STEP2 --> STEP1 STEP2 --> OUT

Approaches can be divided into three primary classes:

  • Geometric-constraint-based methods
  • Feature-similarity-based methods
  • Global-information-based methods

Deep-learning techniques are useful for:

  • generating more robust feature descriptions than conventional methods
  • estimating transformation parameters by embedding to the network

However, lack of training datasets limits the usage in large outdoor scenarios.

2nd: Spatial difference estimation

According to the estimation metric, there are three types of methods:

  • Point-based difference estimation
  • Voxel- or occupancy-grid-based difference estimation
  • Segment- or object-based difference estimation
3rd: Geometric and attribute analysis

Two properties should be derived at this stage:

  • Geometric property: change type, uncertainties, …
  • Attribute property: semantics, …

Discussion

Gaps remain between techniques and demands
  • Dataset reliability: The complexity brought by pseudo-changes is noteworthy.
    • Pseudo-change: changes in non-investigated objects
      • seasonal changes of vegetation covers the surface of buildings
      • scaffolds during construction
      • …
    • TODOs: remove the influence of pseudo-changes
      • shape completion methods
      • filtering methods
      • …
  • Result uncertainty: Occlusion is troublesome.
  • Contributions of semantics: Geometric changes cannot present all possible changes in an observed scene.
Some directions are possible to fill up the gaps
  • Multi-source point clouds
  • Object-level semantics
  • Collaboration with Computer Vision