A Multisensor Dataset for SLAM in Repetitive Scene Environments

초록

Accurate simultaneous localization and mapping (SLAM) is fundamental to autonomous navigation, as it requires both estimating the robot's motion and consistently constructing or aligning a map. However, most existing datasets are collected in feature-rich environments and do not adequately address repetitive scenes that cause perceptual aliasing, drift in odometry, and failures in loop closure. We present a new multisensor dataset specifically designed to evaluate SLAM performance in repetitive scene environments. It includes two representative scenarios: a riverside bikepath that exhibits frame-level repetition and an urban development district that presents block-level repetition. Additional campus sequences are provided as nonrepetitive baselines. The platform integrates two 3-D LiDARs, an red-green-blue-depth (RGB-D) camera, three IMUs, and GNSS. By explicitly incorporating both frame-level and block-level repetitive patterns, this dataset enables systematic analysis of perceptual aliasing in SLAM and serves as a reproducible benchmark for developing robust SLAM systems.

키워드