Source-linked AI summary
UrbanLoco: A Full Sensor Suite Dataset for Mapping and Localization in Urban Scenes
Weisong Wen, Yiyang Zhou, Guohao Zhang, Saman Fahandezh-Saadi, Xiwei Bai, Wei Zhan, Masayoshi Tomizuka, Li-Ta Hsu
TL;DR
Urban mapping and localization remain difficult because urban structures and dynamic objects degrade GNSS, LIDAR, and camera-based methods, while IMUs drift. UrbanLoco addresses this gap with a full sensor-suite dataset spanning challenging urban trajectories and evaluates existing methods in those conditions, where tested SLAM methods perform poorly.
Problem
Dense urban scenes challenge mapping and localization because GNSS suffers signal blockage and multipath, LIDAR and camera methods encounter dynamic objects, and IMUs drift.
Method
UrbanLoco constructs a full sensor-suite mapping and localization dataset with LIDAR, cameras, IMU, and GNSS across challenging urban trajectories.
Results
Tested open-source LIDAR- and camera-based SLAM methods performed poorly in highly urbanized scenarios, with dynamic objects contributing substantially to failures.
Takeaways & Limitations
UrbanLoco provides challenging urban trajectories and multiple sensing modalities for studying mapping and localization under dense urban conditions.
Abstract
from arXiv · showhide
Mapping and localization is a critical module of autonomous driving, and significant achievements have been reached in this field. Beyond Global Navigation Satellite System (GNSS), research in point cloud registration, visual feature matching, and inertia navigation has greatly enhanced the accuracy and robustness of mapping and localization in different scenarios. However, highly urbanized scenes are still challenging: LIDAR- and camera-based methods perform poorly with numerous dynamic objects; the GNSS-based solutions experience signal loss and multipath problems; the inertia measurement units (IMU) suffer from drifting. Unfortunately, current public datasets either do not adequately address this urban challenge or do not provide enough sensor information related to mapping and localization. Here we present UrbanLoco: a mapping/localization dataset collected in highly-urbanized environments with a full sensor-suite. The dataset includes 13 trajectories collected in San Francisco and Hong Kong, covering a total length of over 40 kilometers. Our dataset includes a wide variety of urban terrains: urban canyons, bridges, tunnels, sharp turns, etc. More importantly, our dataset includes information from LIDAR, cameras, IMU, and GNSS receivers. Now the dataset is publicly available through the link in the footnote. Dataset Link: https://advdataset2019.wixsite.com/urbanloco.
1. Introduction
UrbanLoco addresses the lack of public datasets targeting mapping and localization in dense urban environments by providing a full sensor-suite dataset collected across challenging city landscapes.
- Motivation: Few existing public datasets address urban mapping and localization with a cost-effective sensor combination.The dataset comparison is presented alongside the public dataset resource context.
- Challenges: Dense urban mapping and localization are difficult because high-rise structures, dynamic objects, GNSS limitations, and IMU drift undermine existing sensing approaches.GNSS suffers limited satellite visibility and multipath problems, LIDAR and visual methods face inaccurate registration from dynamic objects, and IMUs drift over time.
- Dataset: UrbanLoco provides mapping and localization data from LIDAR, cameras, IMU, and GNSS collected in populous districts of Hong Kong and San Francisco.The dataset combines four essential sensor types in densely populated landscapes.
- Dataset: Over 40 kilometers of trajectories cover urban canyons, bridges, hills, tunnels, and scenes containing numerous dynamic objects.These trajectories target varied and difficult urban driving conditions.
- Sensors: The dataset provides a full sensor suite comprising 1 LIDAR, 6 cameras with 360 degrees view, 1 IMU, and 1 GNSS, with one camera in Hong Kong.Calibration logs are available online.
2. Related Works
Prior mapping and localization datasets and methods provide useful benchmarks and sensor modalities, but their urban coverage, trajectory structure, or sensor completeness remains limited.
- Datasets: KITTI offers long odometry trajectories but was collected in rural, lightly trafficked areas and lacks GNSS and IMU outputs.Its trajectories are generally longer than 500 m, but the scene structure does not adequately challenge dense urban localization.
- Datasets: Oxford RobotCar adds IMU and GPS measurements but was collected in less-urbanized Oxford, limiting its coverage of urban mapping and localization problems.Its additional measurements support testing of more algorithm designs.
- Datasets: NuScenes captures dense traffic and high-rise structures with visual information and IMU, but its 20-second scene segmentation makes long mapping trajectories difficult to obtain.NuScenes was primarily designed for perception and tracking.
- Localization: GNSS-based localization is convenient where reception and satellite visibility are satisfactory, but urban conditions often require expensive GNSS-IMU equipment.Urban environments can cause signal loss and reduce the practicality of such equipment for mass applications.
- Methods: Visual methods use stereo, RGB-D, or camera-IMU fusion, while laser methods rely on point-cloud registration algorithms such as ICP, NDT, and Mi-NoM.The paper evaluates LOAM and NDT on the UrbanLoco dataset.
3. Urbanization Measure
The paper defines urbanization for localization using Skymasks and summarizes each mask with elevation statistics that reflect surrounding building structure.
- Motivation: Urbanization is quantified because high-rise structures block satellites, create multipath and NLOS reception, and often accompany dense traffic.The measure is based on building structures near the ego vehicle.
- Skymask: A Skymask is a polar plot of building silhouettes in which blocked sky is masked and clear overhead sky remains visible.It can be generated from 3D building models or a vehicle-mounted fisheye camera with image segmentation.
- Parameters: The metric uses mean mask elevation angle µMEA and mask elevation angle standard deviation σ^2_MEA to summarize a Skymask.These parameters quantify the elevation distribution of surrounding structures.
- Parameters: N=360 equally spaced azimuth angles are typically used, giving the Skymask a 1-degree azimuth resolution.Each θ_α represents the elevation angle at a given azimuth, which is correlated with building heights.
- Interpretation: Dense urban areas generally produce large µMEA and relatively small σ^2_MEA, whereas rural areas produce smaller values for both parameters.Mixed high-rise and low-lying buildings affect the standard-deviation term.
4. The UrbanLoco Dataset
UrbanLoco combines challenging urban trajectories with a full sensor suite and calibrated collection platforms for mapping and localization. Its Hong Kong and California platforms use different vehicle and sensor configurations, with ground-truth localization from a GNSS-IMU system.
- The dataset covers challenging urban trajectories with a full sensor suite for mapping and localization.
- Hong Kong data were collected using a Honda Fit, while California data were collected using a Toyota Prius.
- The Hong Kong platform includes one LIDAR, a fish-eye camera, an IMU, and a GNSS receiver.
- The California platform includes one LIDAR, six cameras, an IMU, and a GNSS receiver, with the cameras synchronized and calibrated with the LIDAR.
- Ground truth is provided by a Novatel SPAN-CPT GNSS-IMU navigation system, with RTK correction and continuous trajectory output during complete GNSS signal loss.
5. Evaluation
UrbanLoco evaluates state-of-the-art LIDAR- and camera-based SLAM methods in highly urbanized scenes, where dynamic objects, altitude changes, sharp maneuvers, and lighting transitions expose substantial weaknesses. The dataset is presented as a challenging test field for urban-focused localization solutions.
- Evaluation: State-of-the-art SLAM algorithms performed poorly on UrbanLoco’s highly urbanized scenarios, with dynamic objects contributing substantially to failures.Most tested methods use static feature points to estimate vehicle pose, making moving scenes particularly challenging.
- Laser-based Methods: Above 10m/km horizontal translation error occurred for LOAM in scenes filled with dynamic objects.LOAM’s KITTI-reported average translation error was less than 5.7m/km, but performance deteriorated on UrbanLoco.
- Laser-based Methods: Over 20m/km translation error occurred for LOAM under drastic altitude changes, while rotation degraded during sharp maneuvers.These cases correspond to the trajectory conditions illustrated in Figure 4(b,c,d).
- Laser-based Methods: More than 100 meters of overall translation error occurred for NDT on the longest 5.9km testing route.The NDT-based method also performed poorly in altitude estimation despite parameter fine-tuning and half-real-time playback.
- Vision-based Methods: VINS-MONO was highly sensitive to lighting changes and failed when entering or exiting a tunnel.A half-real-time playback rate was used to generate the continuous trajectory successfully.
- Evaluation: The authors conclude that UrbanLoco adequately addresses urban driving challenges and provides a valuable test field for future urban-focused localization solutions.This conclusion follows the dataset’s evaluation of mapping and localization methods in Hong Kong and San Francisco.
6. Conclusion
UrbanLoco is a challenging full sensor-suite dataset for autonomous-vehicle mapping and localization. It combines difficult scenes, poor performance from several state-of-the-art algorithms, an urbanization measure, and LIDAR, cameras, IMU, and GNSS receivers.
- Conclusion: UrbanLoco presents challenging trajectories for autonomous-vehicle mapping and localization, with several state-of-the-art algorithms performing poorly on the dataset.The dataset also defines an urbanization measure to quantify challenges across scenarios.
- Conclusion: The sensor suite includes LIDAR, cameras, IMU, and GNSS receivers for mapping and localization purposes.The authors state that additional data and KITTI-like benchmarks may be added in the future.