Core Concepts
CARLA-Loc is a synthetic dataset designed to facilitate the evaluation of SLAM algorithms under challenging environmental conditions, including diverse weather and dynamic scenarios, by providing a comprehensive sensor setup and ensuring consistent ego motion across sequences.
Abstract
The CARLA-Loc dataset is designed to enable the evaluation of SLAM algorithms under diverse environmental conditions, including varying weather and dynamic scenarios. The dataset is created using the CARLA simulator and integrates a variety of sensors, such as cameras, event cameras, LiDAR, radar, and IMU, with finely tuned parameters to emulate real-world sensor configurations.
The key features of the CARLA-Loc dataset include:
Diverse environmental conditions: The dataset covers 7 different maps with 42 sequences, each featuring varying weather (clear noon, foggy noon, rainy night) and dynamic conditions (static and dynamic).
Comprehensive sensor setup: The dataset provides data from multiple sensors, including stereo RGB cameras, depth cameras, semantic segmentation cameras, event cameras, 32-channel LiDAR, radar, and IMU, with realistic noise and distortion models.
Consistent ego motion: The dataset ensures identical ego trajectories across different environmental settings within each map, enabling direct comparisons of SLAM algorithm performance.
User-friendly pipeline: The dataset comes with a complete pipeline that allows users to record their own sequences according to specific design requirements, using the provided finely tuned and modified sensors. Additionally, a script for converting raw data to ROS bags is included for convenient and efficient data handling.
The authors evaluated several state-of-the-art visual-based and LiDAR-based SLAM algorithms on the CARLA-Loc dataset, analyzing the impact of various challenging environmental factors on localization accuracy. The results demonstrate the utility of the CARLA-Loc dataset in validating the robustness of SLAM algorithms under diverse conditions.
Stats
The dataset provides the following key metrics:
Absolute Position Error (APE) of visual SLAM methods (ORB-SLAM3, VINS-Fusion, S-MSCKF) across different weather and dynamic conditions
Absolute Position Error (APE) of LiDAR SLAM methods (LOAM, LeGo-LOAM, FAST-LIO2) across different weather and dynamic conditions
Quotes
"The robustness of SLAM (Simultaneous Localization and Mapping) algorithms under challenging environmental conditions is critical for the success of autonomous driving."
"To address this, we introduce CARLA-Loc, a synthetic dataset designed for challenging and dynamic environments, created using the CARLA simulator."
"CARLA-Loc comprises 7 maps and 42 sequences, each varying in dynamics and weather conditions."