Study of subsampling strategies in place images matching

Project Overview

This research focuses on enhancing Visual Place Recognition (VPR), a crucial component in robotic SLAM (Simultaneous Localization and Mapping) systems. VPR is essential for Loop Closure Detection, which helps eliminate accumulated errors in SLAM applications.

Problem Definition

The project addresses three key aspects:

  • Database Creation: Managing a set of places {P₁, P₂, …, Pₙ} with corresponding images
  • Query Image Processing: Identifying the location of a query image q within the database
  • Similarity Measurement: Minimizing the distance d(Iᵢ, q) between database and query images

Research Gap & Innovation

Traditional approaches often face limitations:

  • Insufficient attention to discriminative features
  • Over-emphasis on repetitive elements (e.g., sky regions)

Our solution introduces:

  • Novel local descriptor subsampling strategies
  • Adaptive threshold implementation
  • Enhanced attention to discriminative image features

Key Achievements

  • Developed improved subsampling strategies for local descriptors
  • Implemented an adaptive threshold system
  • Achieved superior accuracy compared to the original attentive-patch method
  • Maintained computational efficiency without speed sacrifice

Technical Approach

The system utilizes a multi-stage process:

  1. Initial anchor point selection
  2. Patch creation around selected anchors
  3. Secondary filtering with adaptive thresholding
  4. Final feature set generation

Results

The implemented improvements demonstrated:

  • Enhanced accuracy over the baseline attentive-patch approach
  • Maintained computational efficiency
  • Robust performance in real-world scenarios

Supervisor

Prof. Lam Siew Kei, Nanyang Technological University