Indoor visual SLAM dataset with various acquisition modalities.

Area of Science:

• Robotics
• Computer Vision
• Data Science

Background:

• Visual Simultaneous Localization and Mapping (V-SLAM) is crucial for autonomous systems.
• Evaluating V-SLAM performance requires diverse datasets capturing various environmental conditions and sensor modalities.
• Existing datasets may not fully represent the challenges posed by different acquisition modalities.

Purpose of the Study:

• To create and present a novel indoor V-SLAM dataset.
• To assess the influence of different acquisition modalities (RGB, IR, depth) on V-SLAM accuracy.
• To provide a benchmark for V-SLAM algorithm development and evaluation.

Main Methods:

• Collected data using an Intel RealSense D435i camera with intrinsic calibration.
• Acquired sequences using multiple modalities: RGB, IR, and depth images (passive and active stereo).
• Generated reference trajectories using Structure From Motion (SFM) and Multi View Stereo (MVS) for ground truth.

Main Results:

• The dataset includes spatially aligned and rectified RGB/IR and depth data.
• Sequences feature varied environments: low brightness, changing illumination, wide/narrow spaces, and diverse textures.
• The dataset enables direct comparison of V-SLAM performance across different sensor inputs.

Conclusions:

• The developed dataset serves as a valuable resource for V-SLAM research.
• It facilitates a deeper understanding of how acquisition modalities affect V-SLAM system robustness and accuracy.
• This work contributes to advancing the development of more reliable autonomous navigation systems.