TEX-CUP: The University of Texas Challenge for Urban Positioning
Lakshay Narula, Daniel M. LaChapelle, Matthew J. Murrian, J. Michael Wooten, Todd E. Humphreys, Elliot de Toldi, Guirec Morvant, Jean-Baptiste Lacambre
TTEX-CUP:The University of Texas Challenge for Urban Positioning
Lakshay Narula, Daniel M. LaChapelle, Matthew J. Murrian, J. Michael Wooten, Todd E. Humphreys
Radionavigation LaboratoryThe University of Texas at Austin
Austin, TX, USA
Elliot de Toldi, Guirec Morvant, Jean-Baptiste Lacambre iXblue INC
Denver, CO, USA
Abstract —A public benchmark dataset collected in the denseurban center of the city of Austin, TX is introduced for evaluationof multi-sensor GNSS-based urban positioning. Existing publicdatasets on localization and/or odometry evaluation are basedon sensors such as lidar, cameras, and radar. The role ofGNSS in these datasets is typically limited to the generationof a reference trajectory in conjunction with a high-end inertialnavigation system (INS). In contrast, the dataset introduced inthis paper provides raw ADC output of wideband intermediatefrequency (IF) GNSS data along with tightly synchronized rawmeasurements from inertial measurement units (IMUs) and astereoscopic camera unit. This dataset will enable optimizationof the full GNSS stack from signal tracking to state estimation,as well as sensor fusion with other automotive sensors. Thedataset is available at http://radionavlab.ae.utexas.edu underPublic Datasets. Efforts to collect and share similar datasets froma number of dense urban centers around the world are underway.
Index Terms —urban positioning; precise positioning; bench-mark; dataset; sensor fusion.
I. I
NTRODUCTION
Development of automated ground vehicles (AGVs) hasspurred research in lane-keeping assist systems, automatedintersection management [1], tight-formation platooning, andcooperative sensing [2], [3], all of which demand accurate(e.g., 50-cm at 95%) ground vehicle positioning in an urbanenvironment. But the majority of positioning techniques, andthe associated performance benchmarks, developed thus farare based on lidar or cameras, which perform poorly in low-visibility conditions such as snowy whiteout, dense fog, orheavy rain. Adoption of AGVs in many parts of the worldwill require all-weather localization techniques.Radio-wave-based sensing techniques such as radar andGNSS remain operable even in extreme weather condi-tions [4] because their longer-wavelength electromagnetic radi-ation penetrates snow, fog, and rain. Carrier-phase-differentialGNSS (CDGNSS), also known as real time kinematic (RTK)GNSS, has been successfully applied for the past two decadesas an all-weather decimeter-accurate localization techniquein open-sky conditions. Similarly, inertial sensing techniquesare also unaffected by weather conditions. A combination oflow-cost inertial- and radio-based localization is a promising direction towards precise all-weather urban positioning forAGVs.While application of CDGNSS/RTK techniques for urbanpositioning has previously been limited due to expensive cou-pling with tactical grade IMUs [5], recent work has shown that20-cm-accurate (95%) low-cost unaided CDGNSS positioningis possible at 87% availability with dual-frequency GPS andGalileo signals, even in the dense urban downtown of Austin,TX [6]. Similarly, [7] shows that unaided dual-frequency GPS-, BeiDou-, and GLONASS-based CDGNSS positioning canachieve decimeter-accurate (95%) positioning rate of 76.7%on a 1-hour drive along an urban route in Wuhan, China,and the availability can be further improved to 86.1% afterintegration with a MEMS IMU. Meanwhile, recent urbanCDGNSS evaluation of commercial receivers in [8] indicatesthat no low-to-mid-range consumer CDGNSS solution offersgreater than 35% decimeter-accurate solution availability inurban areas, despite a dense reference network and dual-frequency capability.Similarly, until recently precise point positioning (PPP)algorithms required a long convergence time and as such werelimited to surveying applications. With the proliferation ofthe number of GNSS satellites and better numerical modelsfor atmospheric corrections [9], recent efforts have reportedinstantaneous convergence times for PPP in open sky or lighturban conditions. The authors predict that efforts towardsaccuracy and availability of PPP in urban areas will be soonforthcoming.The concern with development of GNSS-based precisepositioning techniques as described above is that differentalgorithms may have been evaluated on datasets of differentdifficulties, even if the general environment may be describedas urban. As an extreme example, consider the data collectionroute presented in Fig. 3. Over the entirety of the dataset, aSeptentrio AsteRx4 RTK receiver used as a part of this datacollection reports an integer-ambiguity-fixed RTK solution at70.6% of all epochs. However, as is typical, the data collectionroutine involved ≈
10 min stationary open-sky periods at thebeginning and end of data collection. Excluding these periodsbrings down the fixed solution availability to 49.6% on the
Copyright © 2020 by Lakshay Narula May 2020 preprint of paper accepted for publication a r X i v : . [ c s . R O ] M a y emainder of the dataset. In fact, when restricted strictly to thedense urban southern portion of the test route, the availabilityof reported precise RTK solutions is only 21.3%. As such, itis at best challenging, and at worst misleading, to compareprecise GNSS positioning algorithms on different datasets.Additionally, multipath properties of the GNSS antenna andphase stability of the sampling clock are other important fac-tors that likely affect the performance analysis. In the opinionof the authors, the precise GNSS positioning community mustconverge on a shared and challenging dataset to evaluate theiralgorithms and thereby identify the critical components of arobust and accurate urban positioning engine.As a precedent, similar benchmarks such as the KITTIdataset [10] for visual odometry and object segmentation,and the ImageNet dataset [11] for object instance recognitionhave served greatly towards the progress of their respectivecommunities. The Oxford Robotcar Dataset [12] has beena similarly important benchmark dataset in the field of re-peatable ground vehicle localization with lidars and cameras.However, none of the existing robotic localization datasetsare focused on GNSS-based precise urban localization. Thedataset being introduced in this paper addresses this gap forthe GNSS research community.The goal of the University of Texas Challenge for UrbanPositioning is twofold: to enable the precise GNSS positioningcommunity to evaluate and compare a variety of existing andupcoming techniques on a shared and challenging benchmark,and to save the time and effort required to assemble a high-quality data recording platform for urban positioning research.II. S ENSOR P LATFORM
The roving dataset is captured with an integrated perceptionplatform named the University of Texas
Sensorium , shown inFigs. 1 and 2, equipped with the following sensors: • × Antcom G8Ant-3A4TNB1 high performance GNSSpatch antennas (NGS code: ACCG8ANT 3A4TB1).Triple frequency L1/L2/L5;
40 dB low-noise amplifier. • × RadioLynx GNSS RF front end. Dual frequencyL1/L2; bandwidth on both channels; supportfor two GNSS antennas; developed in-house; providedwith Bliley LP-62 low-power
10 MHz
OCXO externalreference. • × NTLab B1065U1-12-X configurable RF front end.Configured to capture L1/L2/L5 signals from one GNSSantenna with a wide bandwidth of
53 Msps ; providedwith Bliley LP-62 low-power
10 MHz
OCXO externalreference. • × u-blox EVK-M8T. Single-frequency (L1) multi-constellation mass-market receiver. • × Bosch BMX055 9-axis IMU. Low-cost MEMSdevice; smartphone-grade IMU noise characteristics;
150 Hz output rate. • × LORD MicroStrain 3DM-GX5-25 AHRS. High-performance MEMS device; industrial-grade IMU noisecharacteristics;
100 Hz output rate. • × Basler acA2040-35gm cameras. × res-olution; monochromatic; Sony IMX265 CMOS sensor;global shutter; hardware triggered at
10 fps ; ≈
50 cm baseline; Kowa LMVZ4411 lenses. • × Delphi ESR 2.5 (24VDC) L2C0051TR electroni-cally scanning radar. Simultaneous mid- and long-rangemeasurement modes; mid-range
60 m , ° field-of-view;long-range
174 m , ° field-of-view;
20 Hz scan rate. • × Delphi SRR2 single beam monopulse radars. Range
80 m ; field-of-view °;
20 Hz scan rate. When mountedas shown in Figs. 1 and 2, the three radars provide °of coverage around the vehicle. • × Taoglas 4G LTE MIMO antenna. Provides connectiv-ity to the network for CDGNSS corrections.For the purposes of this evaluation dataset, the Sensoriumis equipped with an iXblue ATLANS-C: a high-performanceRTK-GNSS coupled fiber-optic gyroscope (FOG) INS (notshown in Figs. 1 and 2). The Septentrio AsteRx4 RTK receiverinside the ATLANS-C is attached to one of the two GNSSantennas, and tracks most constellations on all three GNSSfrequencies. The post-processed fused RTK-INS position so-lution obtained from the ATLANS-C is taken to be the groundtruth trajectory.
LTE Antenna Triple-FrequencyGNSS AntennasDelphi ESR 2.5 RadarDelphi SRR2 Radars Basler acA2040-35gmCameras
Fig. 1. The University of Texas Sensorium is a platform for automatedand connected vehicle perception research. The Sensorium features twoL1/L2/L5 GNSS antennas, wideband GNSS RF front ends, smartphone- andindustrial-grade MEMS IMUs, stereoscopic cameras, automotive radars, andLTE connectivity.
The Sensorium houses a rugged Nuvis N5306RT computerwith a modest desktop-level configuration. The computer runsUbuntu Linux and logs data from all sensors and devices.Most data logging processes are developed in-house for pre-cise synchronization between sensor data. Details on sensorsynchronization are provided in Sec. IV-C.To enable CDGNSS-based positioning, the dataset alsoincludes GNSS data logged from a nearby reference antennawith a clear view of the sky. The reference antenna is ageodetic-grade Trimble Zephyr II (NGS code: TRM57971.00).For consistency with the rover, raw IF reference data is loggedwith identical RadioLynx and NTLab RF front ends. Forcompleteness, RINEX-format reference data from an identicalSeptentrio AsteRx4 receiver is also logged. The rover platform2 ower Strip Nuvis 5306RTMachine Vision ControllerBasler acA2040-35gmCamerasUBlox EVK-M8TGNSS ReceiverNTLab NT1065GNSS Frontend RNL RadiolynxGNSS Frontendwith OCXO Lord MicroStrain3DM-GX5-25 IMU SplittersBias Tee
Fig. 2. Inside view of the University of Texas Sensorium, showing the internalorganization of a desktop-class computer, IMUs, two GNSS RF front ends,and a stereoscopic camera setup. is always within of the reference antenna, representingideal CDGNSS conditions.III. D
ATA C OLLECTION
The test route, depicted in Fig. 3, runs the gamut of light-to-dense urban conditions, from open-sky to narrow streets withoverhanging trees to the high-rise urban city center.The data capture begins and ends with a stationary intervalof several minutes in open sky conditions to allow confidentbookending for the ground truth system. The first part of thetrajectory runs through the semi-urban conditions north of theUniversity of Texas campus, passing under two pedestrianbridges. The second part of the trajectory passes through anarea with narrow streets lined by tall residential apartmentbuildings and dense foliage. The rest of the test route combsthrough the dense urban center of the city of Austin, TX,driving through every east-west street in the city downtown.The number of signals tracked by a receiver is a goodindicator of the level of difficulty posed by the dataset. Fig. 4shows two extremes of this metric by comparing the low-cost mass-market u-blox M8T and the high-performance all-in-view Septentrio AsteRx4. As mentioned before, the u-blox M8T is a single frequency receiver, and is only ableto track GPS and GLONASS signals in the presented dataset.The number of tracked signals during the
30 min challengingdowntown portion of the dataset is under for the M8T,making it unlikely to produce reliable CDGNSS positionestimates [6]. At the other end of the performance spectrum,the AsteRx4 receiver is a state-of-the-art all-in-view receiver,tracking all constellations in all GNSS bands. For this receiver,the number of tracked signals is above for most of thechallenging portion of the dataset.Fig. 5 shows Google Street View imagery from the drivenroute for a qualitative assessment of the dataset difficulty. AKML file with the full route is provided along with the datasetfor easy visualization of the urban conditions.The trajectory shown in Fig. 3 is driven twice, once onThursday, May 9, 2019, and again on Sunday, May 12, Fig. 3. Test route through The University of Texas west campus and Austindowntown. These areas are the most challenging for precise GNSS-basedpositioning. The route was driven once on a weekday and again on theweekend to evaluate robustness of mapping-based methods to changes intraffic and parking patterns. T r ac k e d S i gn a l s ( u - b l ox ) GPS Time of Week (s) T r ac k e d S i gn a l s ( A s t e R x4 ) Fig. 4. A comparison of the number of tracked signals over the durationof the dataset for the u-blox M8T receiver (top) and the Septentrio AsteRx4receiver (bottom). The u-blox M8T is an L1-only receiver tracking GPS andGLONASS signals. The Septentrio AsteRx4 is a triple-frequency all-in-viewreceiver.
A. Data Formats
This section describes the formats of different sensor datamade available as part of this dataset. The description isorganized by the different devices generating the data.
1) RadioLynx Front End:
The RadioLynx RF front endgenerates two-bit-quantized samples from two antennas at therover and a single antenna at the reference station, capturing . bandwidth at both L1 and L2 bands around theGPS frequencies. The raw IF data from the three antennasis made available in a binary format documented along with3 ig. 5. Google Street View imagery of a few challenging scenarios encountered in the dataset. the dataset, including the required IF parameters. Raw IFdata enable development of new signal tracking strategies forurban precise positioning, and allow high-sensitivity receiversto track weak signals that may not have been tracked by thereceivers in the recording platform. Raw IF samples from thereference antenna can be used for data bit wipeoff [6, Sec.III-D], if desired.The dataset also provides tracked pseudorange and carrier-phase observables generated by the GRID software-definedreceiver [6] operating on the RadioLynx raw IF samples forboth rover antennas and the reference antenna. At the time ofwriting, the GRID receiver tracks GPS, Galileo, and SBASsignals. The observables are provided in the RINEX format.
2) NTLab Front End:
The NTLab RF front end producestwo-bit-quantized samples from one of the two rover antennasand the antenna at the reference station. With a sample rate of . , the NTLab front end captures signals at L1, L2, andL5 frequencies with a wide bandwidth. The raw IF data fromboth the rover and reference antennas are made available atthis time, tracked observables in RINEX format will be madeavailable soon.
3) Septentrio AsteRx4:
The Septentrio AsteRx4 receiverhoused inside iXblue ATLANS-C produces observables forGPS, Galileo, GLONASS, BeiDou, and SBAS at all threeGNSS frequencies. All these observables are made availablein RINEX format.
4) u-blox M8T:
The NMEA output from the u-blox M8Treceiver is provided with the dataset for comparison to a competitive mass-market receiver.
5) Stereo Cameras:
Timestamped stereo images from thetwo Basler cameras are made available in HDF5 format. Asdetailed later in Sec. IV-C, camera images are timestampedby the Sensorium computer when the image is received overEthernet. The dataset also provides the exposure time forindividual images if it may be desirable to account for itsvariation.Accurate intrinsic and extrinsic calibration of cameras isimportant for camera-based positioning. This dataset providesan HDF5 archive of stereo and monocular calibration imagescaptured with the Sensorium before the data capture, alongwith measurements of the calibration patterns. These archivesmay be used to obtain both intrinsic and extrinsic calibrationparameters as required, e.g., using the Kalibr calibrationtoolbox [13].
6) Bosch IMU:
To evaluate the benefit of low-cost inertialaiding in urban areas, the dataset includes timestamped specificforce, angular rate, and temperature measurements from theBosch BMX055 IMU in CSV format. This IMU is built-in tothe RadioLynx board, and has been set up such that the IMUdata timestamps can be traced back to the GNSS RF samplingclock. This enables highly accurate correspondence betweenthe IMU timestamps and GPS time.
7) LORD MicroStrain IMU:
Timestamped specific forceand angular rate measurements from the high-performanceLORD MicroStrain MEMS IMU are made available in CSVformat. The LORD IMU accepts a PPS (pulse per second)4ignal generated by the u-blox receiver to synchronize to GPStime. LORD IMU measurements are internally compensatedfor temperature variation.
8) ATLANS-C IMU:
The dataset includes specific force andangular rate measurements from the highly stable accelerom-eters and FOGs housed in the iXblue ATLANS-C. These dataare only made available from the May 9, 2019 data collectionsession. The ATLANS-C data from May 12, 2019 data areheld back for performance evaluation.
9) Ground Truth Trajectory:
A trustworthy ground truthtrajectory against which to compare the reported trajectory of asystem under test is indispensable for urban positioning evalu-ation. Post-processing software provided by iXblue generates aforward-backward smoothed position and orientation solutionwith fusion of AsteRx4 RTK solutions and inertial measure-ments. The post-processed solution is accurate to better than
20 cm throughout the dataset, and may be considered as theground truth trajectory. As with the ATLANS-C IMU measure-ments, the ground truth trajectory is only made available fromthe May 9, 2019 session. The ground truth trajectory from May12, 2019 is withheld for evaluation of community solutions.The authors may advertise the performance of communitysubmissions on the dataset webpage with consent from thedeveloper.
B. Interface with Receivers
The dataset is easiest to interface to with a software-definedreceiver, since these receivers typically accept a stream ofdigitized IF samples as the input. For receivers that onlyaccept RF input, it may be possible to replay the providedraw IF samples after upconversion to RF with use of a GNSSreplay/playback system similar to LabSat 3 Wideband [14].
C. Planned Worldwide Extension
In partnership with iXblue, TEX-CUP is currently beingextended to include raw GNSS IF and IMU data from variousworldwide dense urban centers. These future data captureswill use a simplified version of the Sensorium rover platform,including the same NTLab and RadioLynx front ends, a neweru-blox receiver (ZED-F9P), as well as the Septentrio AsteRx4and ATLANS A7 (upgraded version of the ATLANS-C) orATLANS A9 (best in class) INS for the ground truth trajectory.Raw IMU data from the Bosch BMX055 and ATLANS willalso be included.Urban centers currently under consideration for future datacollection include Denver, CO, Boston, MA, and San Diego,CA in the US, and Paris, Amsterdam, Singapore, and Beijinginternationally.IV. S
ENSOR C ALIBRATION & S
YNCHRONIZATION
Accurate calibration and synchronization of all sensorsis critical for any localization dataset. The performance ofGNSS/INS, odometry, and SLAM techniques strongly dependson the accuracy of sensor calibration and synchronization.
A. Intrinsic Calibration
Intrinsic sensor calibration is necessary for cameras andIMUs, while antenna and front-end calibration may be benefi-cial in high-accuracy and high-availability GNSS applications.
1) Cameras:
Intrinsic camera calibration may be performedby capturing images of a known calibration pattern at differentscales and orientations. The dataset includes such a capturefor the Sensorium cameras. These images may be used with atool such as Kalibr [13] to obtain intrinsic camera parametersincluding focal length, principal point, lens distortion, etc. Itmust be noted that platform vibrations during data collectioncan lead to small variations in the intrinsic calibration param-eters. It is most desirable to continuously track the calibrationparameters in real time in combination with CDGNSS and/orIMUs.
2) IMUs:
The dataset provides a
24 h long stationarycapture of IMU measurements from the Bosch and LORDIMUs to enable calibration of IMU noise and bias stabilityparameters. In addition to noise and bias stability, IMU in-trinsic calibration involves estimation of accelerometer andgyroscope biases and scale factors. Unfortunately, a priori intrinsic calibration is typically not feasible due to variableturn-on-to-turn-on bias properties of the IMUs. It is thuscommon to track the IMU bias and scale factor parameters incombination with GNSS and/or vision-based positioning [15].
3) GNSS Antennas:
Intrinsic calibration of the Sensorium’stwo GNSS antennas amounts to developing a model forantenna phase center variations (PCVs) as a function of thedirection of arrival of an incoming signal. Such a calibrationcan be obtained at the carrier phase level either relative toa reference antenna, as in [16], or in absolute terms, asin [17]. The U.S. National Geodetic Survey (NGS) offersabsolute calibration files for a wide variety of antenna models,including for the type of antenna on the Sensorium (NGS code:ACCG8ANT 3A4TB1) and at the reference station (NGScode: TRM57971.00) . Users of the TEX-CUP data will tendto see improved CDGNSS availability and accuracy whenthese calibrations are applied.PCV models such as offered by the NGS cannot, however,compensate for local effects: The Sensorium’s antennas aremounted on a broad aluminum backplane that affects theantennas’ PCV behavior in a way not captured by the NGSmodel. To obtain a more accurate PCV model for use withthe TEX-CUP data set, a relative calibration was performedbetween each of the Sensorium’s antennas in situ and theTEX-CUP reference antenna. GNSS phase and pseudorangeobservables were collected over a two-day period and aPCV modeling procedure like the one presented in [16]was performed, except that the model was based on double-rather than single-difference carrier phase measurements. Let z represent the zenith angle (angular departure from the antennaboresight), and a represent the azimuth angle of an incomingsignal, both in radians. Then the additional carrier phase ( z, a ) is modeled as d ( z, a ) = (cid:32) n (cid:88) i =1 g i z i (cid:33) m (cid:88) j =1 g cj cos( ja ) + g sj sin( ja ) Azimuthal coefficients g cj and g sj , and elevation coefficients g i , are obtained via a nonlinear least squares fitting procedure.Separate sets of coefficients may be obtained for each of theantennas’ three receiving frequencies.The azimuthal coefficients were found to be too small to beestimated reliably from the two-day data set, but the elevationcoefficients were significant and are available on the TEX-CUP website for both starboard and port Sensorium antennasat both L1 and L2. Application of these coefficients reduces thestandard deviation of L1 and L2 undifferenced carrier phaseresiduals by 11% and 15%, respectively, in the 2-day PCVcalibration data set. Coefficients for the L5 frequency will beposted to the TEX-CUP website in the future.
4) GNSS Code Phase Biases:
Differential code phase bi-ases arise in GNSS receivers due to dissimilar frequency pathsand dissimilar autocorrelation functions [18]. Thus, a biasmay arise between GPS L1 C/A and GPS L2C code phasemeasurements even though the signals have similar autocor-relation properties, and between GPS L1 C/A and Galileo E1measurements even though the signals have identical centerfrequencies. A similar bias exists at each GNSS satellite.Monthly estimates of the satellite-side biases are availablefrom the Center for Orbit Determination in Europe (CODE) .Once these are applied, it is straightforward to estimate thereceiver-side biases relative to a reference signal, usually takento be GPS L1 C/A. During the 10-minute stationary periodsthat bookend each TEX-CUP data interval, a GPS L1 C/A-onlyCDGNSS solution can be obtained for each of the Sensorium’santennas. Due to the short Sensorium-to-reference baseline(less than 1 km during these stationary segments), and toaveraging over the 10-minute period, this solution is accurateto better than 1 cm. Once obtained, this solution can be usedas a truth constraint on the antennas’ location. Next, a high-accuracy ionospheric model such as the final TEC grid of theInternational GNSS Service [19], [20] is applied to compensatefor ionospheric delays in code phase. Finally, the receiver’sdifferential code phase biases are estimated by averagingpseudorange residuals for each signal when the antennas areconstrained to their known location. B. Extrinsic Calibration
Extrinsic calibration involves estimation of the relative po-sitions and orientations of different sensors involved in sensorfusion. Fig. 6 shows the coordinate frames involved in thedataset, and lever arm measurements between these frames.The sensor mounts are machined to maintain ° rotationsbetween the sensors. However, it is possible that the tolerancesinvolved in this process may not be sufficiently accurate forhigh-precision positioning. Accordingly, the provided extrinsic See http://ftp.aiub.unibe.ch/CODE/ parameters should be considered as initial estimates to anonline calibration procedure.Note that the camera calibration data captures mentionedabove may be used to estimate the extrinsic parameters be-tween the two cameras. As noted before, these parametershave been observed to vary due to platform vibrations andmust ideally be tracked in real time.
C. Synchronization
Sensorium IMU and camera measurements are synchro-nized to GPS time. The Bosch IMU is built-in to the Ra-dioLynx board, enabling direct synchronization to the Ra-dioLynx sample clock, and by extension to GPS time. TheLORD MicroStrain IMU accepts a PPS signal from the u-blox receiver and GPS week and whole seconds over USBfrom the software-defined GNSS receiver running on the Sen-sorium computer. The synchronization to GPS time is handledinternally by the LORD IMU. Similarly, the ATLANS-C IMUmeasurements and fused ground truth trajectory are internallysynchronized to GPS time (but reported in UTC time).The Sensorium computer clock is itself synchronized toGPS time to within less than a millisecond by pointing thecomputer’s NTP client to the GPS time reported by thesoftware receiver running on the machine. This enables theSensorium computer to timestamp any sensor data with sub-millisecond accuracy to GPS time.The Basler cameras in the Sensorium accept an externalhardware trigger to capture images. The trigger is generatedby the RadioLynx board at ≈
10 Hz in synchronization withthe sampling clock ticks. As a result, the trigger provided tothe cameras can be traced back to the GNSS sample recordedby the RadioLynx front end. There are two major sources ofdelay that may be taken in to account when processing cameraimages. First, after receiving the hardware trigger, the camerasexpose the sensor for a variable amount of time, depending onthe lighting conditions. Fortunately, the Basler API providesaccess to the exposure time for each image. The provideddataset annotates each individual image with the exposure timereported by the camera. Second, the images are timestampedby the Sensorium computer when these images are receivedover the local Ethernet connection. The data transfer time fromthe camera to the computer is typically very stable since noother devices are on the network, and may be estimated as aconstant parameter in real time, if necessary.V. S
UMMARY & F
UTURE E XTENSIONS
A GNSS-based precise positioning benchmark dataset col-lected in the dense urban center of Austin, TX has been intro-duced. With provision of raw wideband IF GNSS data alongwith tightly synchronized raw measurements from multipleIMUs and a stereoscopic camera unit, the authors hope thatthe precise GNSS positioning community will benefit fromtesting their techniques on a challenging public dataset. Inthe near future, the authors hope to offer a benchmarkingservice similar to the KITTI benchmark suite [10], providingthe opportunity for researchers to publicly compare precise6 xz yx z xyz LordRadiolynx Port CameraStarboard Camera yx z
PRIMARY ALT1iXblue
Fig. 6. Computer drawing showing the position and orientation of the sensors used in this dataset from a top view (top panel) and a front view (bottompanel). Measurements are provided in inches. Note that the indicated coordinate axes are assumed to be perfect ° rotations from each other. The (cid:12) symboldenotes an out-of-page axis, while the ⊗ symbol denotes an in-to-page axis. urban positioning methods. The dataset will soon be extendedto include wideband GNSS IF data collected in several otherurban centers around the world.R EFERENCES[1] D. Fajardo, T.-C. Au, S. Waller, P. Stone, and D. Yang, “Automatedintersection control: Performance of future innovation versus currenttraffic signal control,”
Transportation Research Record: Journal of theTransportation Research Board , no. 2259, pp. 223–232, 2011.[2] J. Choi, V. Va, N. Gonzalez-Prelcic, R. Daniels, C. R. Bhat, and R. W.Heath, “Millimeter-wave vehicular communication to support massiveautomotive sensing,”
IEEE Communications Magazine , vol. 54, no. 12,pp. 160–167, December 2016.[3] D. LaChapelle, T. E. Humphreys, L. Narula, P. A. Iannucci, andE. Moradi-Pari, “Automotive collision risk estimation under cooperativesensing,” in
Proceedings of the IEEE International Conference onAcoustics, Speech, and Signal Processing , Barcelona, Spain, 2020.[4] K. S. Yen, C. Shankwitz, B. Newstrom, T. A. Lasky, and B. Ravani,“Evaluation of the University of Minnesota GPS Snowplow DriverAssistance Program,” Tech. Rep., 2015.[5] M. Petovello, M. Cannon, and G. Lachapelle, “Benefits of using atactical-grade IMU for high-accuracy positioning,”
Navigation, Journalof the Institute of Navigation , vol. 51, no. 1, pp. 1–12, 2004.[6] T. E. Humphreys, L. Narula, and M. J. Murrian, “Deep urban unaidedprecise GNSS vehicle positioning,”
IEEE Intelligent TransportationSystems Magazine , 2020, to be published. [7] T. Li, H. Zhang, Z. Gao, Q. Chen, and X. Niu, “High-accuracypositioning in urban environments using single-frequency multi-GNSSRTK/MEMS-IMU integration,”
Remote Sensing , vol. 10, no. 2, p. 205,2018.[8] J. Jackson, B. Davis, and D. Gebre-Egziabher, “An assessment of low-cost RTK GNSS receivers,” in
Proceedings of the IEEE/ION PLANSMeeting , Monterey, CA, 2018.[9] S. Banville, P. Collins, W. Zhang, and R. B. Langley, “Global and re-gional ionospheric corrections for faster PPP convergence,”
Navigation ,vol. 61, no. 2, pp. 115–124, 2014.[10] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics:The KITTI dataset,”
International Journal of Robotics Research (IJRR) ,2013.[11] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet:A large-scale hierarchical image database,” in . IEEE, 2009, pp. 248–255.[12] W. Maddern, G. Pascoe, C. Linegar, and P. Newman, “1 Year, 1000km:The Oxford RobotCar Dataset,”
The International Journal of RoboticsResearch (IJRR) , vol. 36, no. 1, pp. 3–15, 2017. [Online]. Available:http://dx.doi.org/10.1177/0278364916679498[13] P. Furgale, J. Rehder, and R. Siegwart, “Unified temporal and spatialcalibration for multi-sensor systems,” in arXiv preprint arXiv:1704.06053 , 2017.[16] G. L. Mader, “GPS antenna calibration at the National Geodetic Survey,” PS solutions , vol. 3, no. 1, pp. 50–58, 1999.[17] A. Bilich and G. L. Mader, “Gnss absolute antenna calibration atthe national geodetic survey,” in
Proceedings of the 2010 ION GNSSConference , 2010.[18] O. Montenbruck, A. Hauschild, and P. Steigenberger, “Differential codebias estimation using multi-gnss observations and global ionospheremaps,”
Navigation: Journal of the Institute of Navigation , vol. 61, no. 3,pp. 191–201, 2014.[19] M. Hern´andez-Pajares, J. M. Juan, J. Sanz, R. Orus, A. Garcia-Rigo, J. Feltens, A. Komjathy, S. C. Schaer, and A. Krankowski, “The igsvtec maps: a reliable source of ionospheric information since 1998,”
Journal of Geodesy , vol. 83, no. 3, pp. 263–275, Mar 2009. [Online].Available: https://doi.org/10.1007/s00190-008-0266-1[20] L. Narula, J. M. Wooten, M. J. Murrian, D. M. LaChapelle, andT. E. Humphreys, “Accurate collaborative globally-referenced digitalmapping with standard GNSS,”
Sensors