LiDAR Configuration Comparison for Urban Mapping System Joowan Kim, Jinyong Jeong, Young-Sik Shin, Younggun Cho, Hyunchul Roh and Ayoung Kim 1* 1 Department of Civil and Environmental Engineering, KAIST, Republic of Korea (Tel : +82-042-350-3672; E-mail: [jw kim, jjy0923, youngsik, yg.cho, rohs , ayoungk]@kaist.ac.kr) Abstract—The Mobile Mapping System (MMS) is widely used when mapping urban environment. The critical challenge for mapping accuracy is at localization accuracy under highly sporadic global positioning system (GPS) signal. To tackle this issue, approaches often rely on cameras and Light Detection and Ranging (LiDAR)s to exploit visual and spatial features in the environment. Among many sensors, this paper focuses on the use of LiDAR, especially evaluating the LiDAR types and mechanical configurations. In this paper, we compare two typical LiDAR configurations, push-broom (2D) and 360 scanning (3D) style, in terms of the resulting mapping performance. Resulting maps from two configurations over the same place are directly compared to evaluate characteristic of each LiDAR configuration. Keywords—Mobile Mapping System, LiDAR, SLAM 1. I NTRODUCTION Many mobile mapping systems aim at constructing an accurate map of urban environment by merging navigational and perceptual sensors. Toward this objective, recent mobile mapping systems researches focus on a dense 3D maps from LiDAR and visual recognition. Among many perceptual sen- sors, LiDARs are gaining popularity for many years for dense map generation and point cloud based localization. Typical LiDAR sensors can be categorized as 2D and 3D LiDAR. The 2D LiDAR sensors are array of 1D LiDARs, and each 2D scan produces a slice of point cloud data. These 2D LiDAR are often mounted perpendicular or horizontal to the ground, so the relationship between the acquired point cloud and the vehicle coordinate system can be directly configured. For the 2D LiDAR based systems, it is more straightforward to obtain high-level building information. The 3D LiDAR rotates the array of 1D LiDARs constructing 360 ◦ field of view. In the case of using 3D LiDAR, researches focus on information centering on the building adjacent to the vehicle at the ground level. Since the point cloud is continuously updated even when the vehicle does not move, it has an advantage when handling the occlusion in both static and dynamic environment. Each LiDAR configuration possesses its own advantages and drawbacks. Using 2D LiDAR might be vulnerable to vacancies in occluded area when scanning a slice of mea- surement from the urban structure per measurement. The 3D LiDAR is beneficial being robust against occluded region but needs to deal with large data and motion induced data incon- sistency. This paper addresses this difference and compares the two systems in terms of mapping performance. 2. RELATED WORKS Similar types of urban mapping systems as ours have been introduced in the literature [1, 2, 3, 4]. Blanco et al., [1] developed similar mobile mapping system with vertically installed 2D LiDARs. In their follow-up research, the 2D LiDAR was also used [2]. Pandey et al., [5] used vehicle with both 2D and 3D LiDARs. Their vehicle was equipped with a 3D LiDAR scanner (Velodyne HDL-64E) and two 2D push- broom forward-looking LiDARs (Riegl LMS-Q120). Modified platform appeared in their follow up research [6]. Instead of using both 3D and 2D LiDARs, their vehicles were equipped with four 3D LiDAR scanners (Velodyne HDL-32E). In their work, a prior map was generated by using the LiDAR. Then the vehicle was localized to this prior map using a monocular camera based on mutual information (MI). Similar to our works, system analysis has been reported. Petrie [7] analyzed several mobile mapping systems which were commercially available. In the paper, author presented both imaging and mapping performance of the vehicles, pro- viding detailed survey of widely used platforms. Street Mapper 360 [8] developed a mapping module that is applicable to various mobile platform. The module is capable of producing a rich (e.g., 600,000 points/sec) point cloud even under motion, while providing a cm-level accuracy. Topcon IP-S3 HD1 [9] also provides mapping system with high-quality data. The platform maps the environment using a high-speed scanning module (e.g., 700,000 points/sec). 3. TWO URBAN MAPPING SYSTEMS This paper compares aforementioned two LiDAR configu- rations in terms of map data density, obstructed area, and point cloud richness. The presented two sensor systems commonly include the configuration of the four cameras, inertial measure- ment unit (IMU), fiber optic gyro (FOG), conventional GPS, and the encoder system for vehicle odometry. 3.1. Push Broom 2D Scanning LiDAR First type of mobile mapping system is as shown in 1(a). This system is a car-type vehicle with three 2D LiDARs, four RGB cameras and other navigational sensors. The vehicle pose is estimated by the IMU, Differential Global Positioning System (DGPS), and wheel encoders mounted on the vehicle. Attitude data is collected using IMU that runs in 100 Hz with 0.1 ◦ accuracy. Rotation of each wheel is measured using wheel encoder (100 Hz). For data logging and processing, we use a desktop PC configured with Intel i7 CPU (3.4GHz), 8GB RAM and SSD storage. Two 2D LiDARs are pointing each side of the vehicle with sweeping plane perpendicular to the ground. By mounting the LiDARs in this configuration, we accumulated a slice of data per measurement and achieved data around the road while the vehicle was driving. Based
4
Embed
LiDAR Configuration Comparison for Urban Mapping …irap.kaist.ac.kr/publications/mars-2017-urai.pdf · LiDAR Configuration Comparison for Urban Mapping System Joowan Kim, Jinyong
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
LiDAR Configuration Comparison for Urban Mapping System
Joowan Kim, Jinyong Jeong, Young-Sik Shin, Younggun Cho, Hyunchul Roh and Ayoung Kim1∗1Department of Civil and Environmental Engineering, KAIST, Republic of Korea
Abstract—The Mobile Mapping System (MMS) is widely usedwhen mapping urban environment. The critical challenge formapping accuracy is at localization accuracy under highlysporadic global positioning system (GPS) signal. To tackle thisissue, approaches often rely on cameras and Light Detectionand Ranging (LiDAR)s to exploit visual and spatial features inthe environment. Among many sensors, this paper focuses onthe use of LiDAR, especially evaluating the LiDAR types andmechanical configurations. In this paper, we compare two typicalLiDAR configurations, push-broom (2D) and 360 scanning (3D)style, in terms of the resulting mapping performance. Resultingmaps from two configurations over the same place are directlycompared to evaluate characteristic of each LiDAR configuration.
Keywords—Mobile Mapping System, LiDAR, SLAM
1. INTRODUCTION
Many mobile mapping systems aim at constructing anaccurate map of urban environment by merging navigationaland perceptual sensors. Toward this objective, recent mobilemapping systems researches focus on a dense 3D maps fromLiDAR and visual recognition. Among many perceptual sen-sors, LiDARs are gaining popularity for many years for densemap generation and point cloud based localization.
Typical LiDAR sensors can be categorized as 2D and 3DLiDAR. The 2D LiDAR sensors are array of 1D LiDARs, andeach 2D scan produces a slice of point cloud data. These 2DLiDAR are often mounted perpendicular or horizontal to theground, so the relationship between the acquired point cloudand the vehicle coordinate system can be directly configured.For the 2D LiDAR based systems, it is more straightforwardto obtain high-level building information.
The 3D LiDAR rotates the array of 1D LiDARs constructing360◦ field of view. In the case of using 3D LiDAR, researchesfocus on information centering on the building adjacent tothe vehicle at the ground level. Since the point cloud iscontinuously updated even when the vehicle does not move, ithas an advantage when handling the occlusion in both staticand dynamic environment.
Each LiDAR configuration possesses its own advantagesand drawbacks. Using 2D LiDAR might be vulnerable tovacancies in occluded area when scanning a slice of mea-surement from the urban structure per measurement. The 3DLiDAR is beneficial being robust against occluded region butneeds to deal with large data and motion induced data incon-sistency. This paper addresses this difference and comparesthe two systems in terms of mapping performance.
2. RELATED WORKS
Similar types of urban mapping systems as ours havebeen introduced in the literature [1, 2, 3, 4]. Blanco et al.,
[1] developed similar mobile mapping system with verticallyinstalled 2D LiDARs. In their follow-up research, the 2DLiDAR was also used [2]. Pandey et al., [5] used vehicle withboth 2D and 3D LiDARs. Their vehicle was equipped with a3D LiDAR scanner (Velodyne HDL-64E) and two 2D push-broom forward-looking LiDARs (Riegl LMS-Q120). Modifiedplatform appeared in their follow up research [6]. Instead ofusing both 3D and 2D LiDARs, their vehicles were equippedwith four 3D LiDAR scanners (Velodyne HDL-32E). In theirwork, a prior map was generated by using the LiDAR. Thenthe vehicle was localized to this prior map using a monocularcamera based on mutual information (MI).
Similar to our works, system analysis has been reported.Petrie [7] analyzed several mobile mapping systems whichwere commercially available. In the paper, author presentedboth imaging and mapping performance of the vehicles, pro-viding detailed survey of widely used platforms. Street Mapper360 [8] developed a mapping module that is applicable tovarious mobile platform. The module is capable of producinga rich (e.g., 600,000 points/sec) point cloud even under motion,while providing a cm-level accuracy. Topcon IP-S3 HD1 [9]also provides mapping system with high-quality data. Theplatform maps the environment using a high-speed scanningmodule (e.g., 700,000 points/sec).
3. TWO URBAN MAPPING SYSTEMS
This paper compares aforementioned two LiDAR configu-rations in terms of map data density, obstructed area, and pointcloud richness. The presented two sensor systems commonlyinclude the configuration of the four cameras, inertial measure-ment unit (IMU), fiber optic gyro (FOG), conventional GPS,and the encoder system for vehicle odometry.
3.1. Push Broom 2D Scanning LiDAR
First type of mobile mapping system is as shown in 1(a).This system is a car-type vehicle with three 2D LiDARs,four RGB cameras and other navigational sensors. The vehiclepose is estimated by the IMU, Differential Global PositioningSystem (DGPS), and wheel encoders mounted on the vehicle.Attitude data is collected using IMU that runs in 100 Hz with0.1◦ accuracy. Rotation of each wheel is measured using wheelencoder (100 Hz). For data logging and processing, we usea desktop PC configured with Intel i7 CPU (3.4GHz), 8GBRAM and SSD storage. Two 2D LiDARs are pointing eachside of the vehicle with sweeping plane perpendicular to theground. By mounting the LiDARs in this configuration, weaccumulated a slice of data per measurement and achieveddata around the road while the vehicle was driving. Based
(a) Push broom type 2D LiDAR (b) Tilted 360 type 3D LiDAR
Dimensions 1.67 m × 1.36 m × 0.31 m (L × W × H )Dry weight 35.8 kg
Processor Intel(R) Core(TM) i7-6700 [email protected] Devicemall 100 Ah, 28 V , lithium-iron type
(d) Specifications of UMS sensors
Fig. 1. Sensor configuration for two mobile mapping systems. Below are the summaries of the specification.
on a recommendation [10], the vehicle speed was limitedto 15− 30 km/h. This vehicle speed limitation would bethe critical limitation for mapping when the fast mapping isrequired.
Encoder
IMU
FOG
GPS
VRS-GPS
Altimeter
2D LiDAR
Camera
Odometry
Client
(6D pose)
iSAM
(SLAM
back-end)
3D
World
model
Colored
Point cloud3D LiDAR
Fig. 2. Sensor data diagram for two urban mapping systems. The color of eachletter represents the sensor configuration of the two systems. Black indicatessensors the commonly used in both systems, and red and blue are the sensorsin the 2D LiDAR system and the 3D LiDAR system respectively.
3.2. Tilted 360 Scanning 3D LiDAR
Second type of the platform is equipped with two 3DLiDARs (Velodyne VLP-16) as shown in Fig. 1(b). Similarto the 2D LiDAR, the detailed sensor configuration is shownin the table (Fig. 1(d)). Two tilted LiDARs, four cameras,wheel encoders and GPS are mounted outside of the vehicle.Wheel encoders and IMU are the same specifications aspreviously mentioned systems with 2D LiDARs. The tiltedmount for the two LiDARs minimizes laser shadowing region.The advantage of using this LiDAR configuration is that eachLiDAR compensates another’s occluded area. Through post-processing, angle, range, and intensity data are fused togetherfrom both LiDARs to create a more complete and accurate 3Dmap. The interior of the vehicle is equipped with FOG, IMUand wheel encoder modules. All of these sensors are processedby a desktop PC configured with Intel(R) Core(TM) [email protected] installed inside the vehicle.
3.3. Sensor System Software
For both sensor configurations, our software architecture isas shown in Fig. 2. The 6-degree of freedom (DOF) vehiclepose is calculated using navigational sensors. Using thesecomputed poses, colored point cloud is obtained by fusingLiDAR and camera images. As the simultaneous localizationand mapping (SLAM) back-end, incremental smoothing andmapping (iSAM) is used to optimize entire vehicle trajectory.
Fig. 3. Two sample maps obtained from each mobile mapping system
Accurate 3D world model is reconstructed with respect to thecomputed vehicle trajectory. The more detailed explanationon SLAM application reaches beyond the scope of this paper,and we refer readers to our previous works on map generationusing the aforementioned systems [11, 12, 13].
4. TWO SYSTEM ANALYSIS
In this section, we present the comparison results fromtwo different LiDAR configurations described earlier. Twoexperiments over the same environment were performed byusing each sensor system. For both tests, we used synchro-nized cameras, IMU, wheel encoders and Light Detection andRanging (LiDAR)s. Using SLAM based approach [13], wegenerated 3D map of the environment as shown in Fig. 3.
4.1. Data Density
As shown in Table. 1, overlapping points in the case of atilted 360 scanning system are significantly greater than in thepush broom style mapping system. The 2D LiDAR system hasa 180◦ field of view (FOV) and sensing range is 80 m. On theother hand, the 3D LiDAR of the tilted system has a 360◦ FOVand sensing range is 100 m. The 3D LiDAR system allowsmore vertical objects and wall scanning along the vehicle’sdirection of travel, and hence a richer distribution of points
TABLE 1SUMMARY OF VOXEL CONVERSION RESULT FOR PUSH BROOM
for buildings and trees. Blue box in Fig. 3(a) and Fig. 3(d)shows substantial density difference in the two configurations.
However, the cost is at the data size. The 2D LiDAR config-uration collects measurement data of the defined scan range75 times per second (361 measurement data per scan). The3D LiDAR, on the other hand, takes 1.33 ms to accumulateone data packet [14]. This implies a data rate of 754 datapackets per second. In particular, we can see that the number ofpoints is about three times more than 2D LiDAR system whenanalyzing buildings and trees in the blue box. Our comparisonshows that the appearance of the object is richer with the 3DLiDAR while larger data packet size is required.
4.2. Blind Region
As described in the §3.2, The 3D LiDAR based mappingsystem was mounted on the vehicle at an angle of about 45◦.Therefore, the system avoids occlusion by any close objects(e.g., nearby buildings and cars), enabling the system to have asmaller blind region. The red squares in Fig. 3(a) and Fig. 3(d)show these blind regions caused by occlusion. Correspondingpoint cloud regions are empty when mapped by 2D LiDAR(Fig. 3(b) and Fig. 3(e)). On contrary, the point cloud is farricher when using 3D LiDAR.
5. CONCLUSION
This paper reported comparison between two typical LiDARconfigurations which were widely used in the MMS. Forpush broom scanning system with 2D LiDARs, the LiDARswere with a single channel and a relatively low FOV. Thisconfiguration provided direct and easy implementation but waslimited when creating a dense map with low vehicle speed.Another mapping system was equipped with 3D LiDAR that
has 16 channels with 360◦ scanning ability. The comparisonreported from real experiments showed that the blind area canbe minimized through the tilted 3D LiDAR configuration butwith larger data packet required.
ACKNOWLEDGMENT
This work was supported by [SLAM-based Lane Map Gen-eration for Complex Urban Environment] project funded byNaverLabs Corporation. This material is also based upon worksupported by the MOTIE, Korea under Industrial TechnologyInnovation Program (No. 10067202 and No. 10051867). J.Kim and J. Jeong were supported by MOLIT, Korea via U-City Master and Doctor Course Grant Program.
REFERENCES
[1] J.-L. Blanco, F.-A. Moreno, and J. Gonzalez-Jimenez, “Acollection of outdoor robotic datasets with centimeter-accuracy ground truth,” Autonomous Robots, vol. 27,no. 4, pp. 327–351, Nov 2009.
[2] ——, “The malaga urban dataset: High-rate stereo andLiDARs in a realistic urban scenario,” InternationalJournal of Robotics Research, vol. 33, no. 2, pp. 207–214, 2014.
[3] J. Elseberg, D. Borrmann, and A. Nchter, “Algorithmicsolutions for computing precise maximum likelihood3d point clouds from mobile laser scanning platforms,”Remote Sensing, vol. 5, no. 11, pp. 5871–5906, 2013.
[4] Y. Bok, D.-G. Choi, Y. Jeong, and I. S. Kweon, “Captur-ing city-level scenes with a synchronized camera-laserfusion sensor,” in Proceedings of the IEEE/RSJ Inter-national Conference on Intelligent Robots and Systems.IEEE, 2011, pp. 4436–4441.
[5] G. Pandey, J. R. McBride, and R. M. Eustice, “Ford cam-pus vision and LiDAR data set,” International Journal ofRobotics Research, vol. 30, no. 13, pp. 1543–1552, 2011.
[6] R. W. Wolcott and R. M. Eustice, “Fast LIDAR local-ization using multiresolution Gaussian mixture maps,”in Proceedings of the IEEE International Conference onRobotics and Automation, Seattle, WA, USA, May 2015,pp. 2814–2821.
[7] G. Petrie, “An introduction to the technology mobilemapping systems,” Geoinformatics, pp. 32–43, 2010.
[9] “Topcon ip-s3,” http://www.topcon.co.jp/, TOPCONCORPORATION, Tokyo , Japan, 2015.
[10] J. Borenstein and L. Feng, “Measurement and correctionof systematic odometry errors in mobile robots,” IEEETransactions on Robotics and Automation, vol. 12, no. 6,pp. 869–880, Dec 1996.
[11] J. Jeong, Y. Cho, and A. Kim, “Road-SLAM : Roadmarking based SLAM with lane-level accuracy,” in Pro-ceedings of the IEEE Intelligent Vehicle Symposium,Redondo Beach, CA, Jun. 2017, accepted. To appear.
[12] J. Jeong and A. Kim, “Adaptive inverse perspectivemapping for lane map generation with SLAM,” in Pro-ceedings of the IEEE Ubiquitous Robots and AmbientIntelligence (URAI), Xian, China, Aug. 2016, pp. 38–41.
[13] H. Roh, J. Jeong, Y. Cho, and A. Kim, “Accurate mobileurban mapping via digital map-based SLAM,” MDPISensors, vol. 16, no. 8, p. 1315, Aug. 2016.
[14] “Velodyne VLP-16,” http://velodynelidar.com, VelodyneLiDAR, 5521 Hellyer Ave, San Jose, CA 95138, 2015.