Amirkabir campus dataset: Real-world challenges and scenarios of Visual Inertial Odometry (VIO) for visually impaired people

TL;DR

This paper introduces AUT-VI, a highly challenging Visual Inertial Odometry (VIO) dataset intended to advance navigation tools for visually impaired users. The dataset comprises 126 sequences across 17 campus locations, featuring dynamic objects, diverse lighting, reflections, and abrupt camera motions, and is complemented by an Android data-capture app (VIRec) to enable researcher-driven dataset customization with GPS-ground-truth. The authors provide detailed data formats, calibration procedures, and sequence statistics, and they evaluate leading VO/VIO/SLAM methods (Basalt, VINS-Mono, ORB-SLAM3, SLAMANTIC) to benchmark performance under real-world challenges, including dynamic occlusions and day/night loop-closure scenarios. The work argues that none of the current methods fully address all the challenges captured by AUT-VI, highlighting gaps and guiding future improvements such as dynamic-object segmentation, inertial-only estimation in extreme scenarios, and improved feature matching (e.g., Superglue) for robust loop-closure.

Abstract

Visual Inertial Odometry (VIO) algorithms estimate the accurate camera trajectory by using camera and Inertial Measurement Unit (IMU) sensors. The applications of VIO span a diverse range, including augmented reality and indoor navigation. VIO algorithms hold the potential to facilitate navigation for visually impaired individuals in both indoor and outdoor settings. Nevertheless, state-of-the-art VIO algorithms encounter substantial challenges in dynamic environments, particularly in densely populated corridors. Existing VIO datasets, e.g., ADVIO, typically fail to effectively exploit these challenges. In this paper, we introduce the Amirkabir campus dataset (AUT-VI) to address the mentioned problem and improve the navigation systems. AUT-VI is a novel and super-challenging dataset with 126 diverse sequences in 17 different locations. This dataset contains dynamic objects, challenging loop-closure/map-reuse, different lighting conditions, reflections, and sudden camera movements to cover all extreme navigation scenarios. Moreover, in support of ongoing development efforts, we have released the Android application for data capture to the public. This allows fellow researchers to easily capture their customized VIO dataset variations. In addition, we evaluate state-of-the-art Visual Inertial Odometry (VIO) and Visual Odometry (VO) methods on our dataset, emphasizing the essential need for this challenging dataset.

Figures (4)

• Figure 1: Example scenes from the AUT-VI dataset. A) There are multiple locations (indoors and outdoors) over the campus where the dataset is recorded. B) More than half of the sequences in the dataset are completely dynamic and challenging. C) The videos are captured at various times of the day for each location with almost the same trajectory.
• Figure 2: Details of cell-phone sensors used to record AUT-VI. The setup used to record this dataset has two cameras. The back camera shown at left is responsible for the main images of the dataset. The IMU is located near the cameras, and it is shown at both left and right figures.
• Figure 3: Two example of sequences available inside the dataset. A) A walk around the campus in the evening. B) A walk from center of campus to the bio-medical Engineering department.
• Figure 4: Dataset statistics. A) Histogram distribution of sequence lengths. B) Statistics about the exact environment and time of recording of sequences.