定义：Visual-Based Localization (VBL) consists of retrieving the pose of a visual query material within a known space representation。 论文：A survey on Visual-Based Localization: On the benefit of heterogeneous data

方法分类

  • 间接法(indirect methods)
  • 直接法(direct methods)

两种方法

数据表示

局部特征：特征提取和描述
  • 点特征
  • 几何特征
  • 点&几何特征
全局特征
  • 人工特征
  • 学习特征
混合特征
  • Patch特征(介于局部与全局特征之间)
  • 混合特征
语义特征

间接法VBL

步骤：

  1. Features Aggregation
    • Feature to visual word assignment: DBoW
    • Weighting scheme: tf-idf
    • CNN: NetVLAD
  2. Similarity Research
    • PCA
    • NN: FLANN
    • 机器学习
    • (Z)SSD/SAD/NCC
  3. Condidates re-ranking

直接法VBL

三种方式

  • Direct VBL with prior: pose prior can be obtained through localization sensor (GPS, magnetic compass, 间接法VBL)
  • Features to points matching
    • F2P (Features to Points)
    • PnP
  • Pose regression approaches
    • CNN
    • regression forest

Data with Dissimilar Appearances

  • Appearance changes
    • Viewpoint changes
    • Illumination invariance & long-term localization
    • Dynamic scene
  • Cross-appearance localization
    • Cross-view
    • Cross-domain

异构数据

  • Optical information
  • Geometric information
  • Semantic information

数据融合