PaperlayerView exact revision on arXiv

Source-linked AI summary

I-24 MOTION: An instrument for freeway traffic science

Derek Gloudemans, Yanbing Wang, Junyi Ji, Gergely Zachar, Will Barbour, Daniel B. Work

arXiv:2301.11198v2eess.IVcs.CV

TL;DR

Traffic science needs observations that connect individual vehicle behavior with traffic patterns across space and time, while trajectory data remain limited. I-24 MOTION uses pole-mounted cameras to produce freeway trajectories and publishes initial datasets. The paper reports that its observations already show recurring congestion waves and can support macroscopic analysis, while noting known data artifacts and that the current illustrations are preliminary.

  • Problem

    Traffic science needs long-duration, spatially extensive observations of individual and collective vehicle behavior, but trajectory datasets remain limited.

    Source: §I · p. 1

  • Method

    I-24 MOTION uses overlapping pole-mounted cameras and a fiber-connected compute facility to convert freeway video into vehicle trajectories.

    Source: §I · p. 2

  • Results

    The released trajectories show recurring upstream traffic waves at approximately 12–13 miles per hour across a four-hour, 4.2-mile westbound morning-congestion example.

    Source: §I · p. 2

  • Takeaways & Limitations

    The initial release can support macroscopic analysis without interpolation across the observed four-mile segment, and the instrument is intended to support live traffic experiments.

    Source: §VI · p. 16

    The current illustrations are not comprehensive, and the initial datasets include known issues such as fragmented trajectories from tracking failures and crash-damaged hardware.

    Source: B. Fundamental diagrams (FDs) · p. 15

Abstract

from arXiv · show

The Interstate-24 MObility Technology Interstate Observation Network (I-24 MOTION) is a new instrument for traffic science located near Nashville, Tennessee. I-24 MOTION consists of 276 pole-mounted high-resolution traffic cameras that provide seamless coverage of approximately 4.2 miles I-24, a 4-5 lane (each direction) freeway with frequently observed congestion. The cameras are connected via fiber optic network to a compute facility where vehicle trajectories are extracted from the video imagery using computer vision techniques. Approximately 230 million vehicle miles of travel occur within I-24 MOTION annually. The main output of the instrument are vehicle trajectory datasets that contain the position of each vehicle on the freeway, as well as other supplementary information vehicle dimensions and class. This article describes the design and creation of the instrument, and provides the first publicly available datasets generated from the instrument. The datasets published with this article contains at least 4 hours of vehicle trajectory data for each of 10 days. As the system continues to mature, all trajectory data will be made publicly available at i24motion.org/data.

I. INTRODUCTION

I-24 MOTION addresses the limited availability of freeway trajectory data with a camera-based instrument and initial trajectory datasets. Its continuous observations span individual vehicle trajectories and recurring freeway congestion, supporting traffic analysis across microscopic and macroscopic scales.

A. Data collection for traffic modeling

  • The introduction frames this need in the context of automated vehicles that may alter macroscopic traffic behavior and emerging sensing technologies that capture vehicle context and roadway-scale observations.It emphasizes monitoring individual and collective behavior over long temporal and spatial resolutions.Source: §I · p. 1

I. INTRODUCTION

  • I-24 MOTION responds to limited trajectory-data availability following freeway datasets such as NGSIM and HighD and urban data such as pNEUMA.The authors describe a five-year effort to instrument a freeway section for empirical traffic science.Source: §I · p. 1
  • 276 pole-mounted cameras and a fiber network feed video to a compute facility that converts it into trajectories across a congested freeway segment.The instrument captures approximately 230 million vehicle-miles of travel annually.Source: §I · p. 2
  • The initial release includes trajectories across 4.2 miles and four hours of westbound morning congestion, with vehicle dimensions and lateral positions also measured.Recurring waves in the displayed data propagate at approximately 12–13 miles per hour.Source: Figure 1 · p. 1, §I · p. 2
  • The paper presents the instrument, its infrastructure and processing, initial datasets, and the location for future releases.The article also describes known artifacts and plans to improve the datasets over time.Source: §I · p. 2

A. Data collection for traffic modeling

  • Trajectory data can connect microscopic vehicle interactions with macroscopic traffic waves, which aggregate measurements alone cannot explain in terms of wave generation and propagation.The cited discussion also describes applications to jam clusters, state transitions, and heterogeneous traffic interactions.Source: A. Data collection for traffic modeling · p. 3

B. Existing Testbeds

I-24 MOTION is an open-road testbed designed to provide high-fidelity trajectories across a multi-mile freeway segment in live traffic. Its location, pole-mounted camera network, and infrastructure support continuous observation and future study of corridor technologies.

B. Existing Testbeds

  • I-24 MOTION provides an open-road setting for live-traffic experiments, while closed-course testbeds offer controlled conditions but may not reflect public-road traffic.The paper notes that multi-mile, high-fidelity trajectory collection on every U.S. roadway vehicle was not then available.Source: B. Existing Testbeds · p. 3

C. Emerging Observation Technologies

  • Aerial datasets such as HighD and ExiD capture complete highway trajectories, but drone battery life and human piloting limit observation time.Other drone-based datasets study urban vehicle and pedestrian interactions.Source: C. Emerging Observation Technologies · p. 4

III. SYSTEM DESCRIPTION

  • I-24 MOTION combines freeway infrastructure, network and compute hardware, and algorithms to produce vehicle trajectories at large spatial and temporal scales.The system remains in active development, with planned improvements to reliability, accuracy, and processing speed.Source: §III · p. 4

A. Physical Infrastructure

  • The instrument covers 4.2 miles of I-24 using 276 cameras on 40 tall poles, whose overhead views and overlapping fields of view reduce occlusion.The poles are spaced every 500–600 feet and connect by fiber to a data center for computer-vision tracking and trajectory processing.Source: A. Physical Infrastructure · p. 4

1) Location:

  • The Nashville corridor was selected for recurring congestion, construction feasibility, and co-location with TDOT initiatives.The site carries about 150,000 vehicles daily, including 10–15% heavy trucks; the I-24 SMART Corridor may eventually support study of its traffic-management solutions.Source: 1) Location · p. 4

2) Camera poles:

  • The camera-pole system was prototyped over three years, and lessons from a three-pole validation system informed camera selection, mounting, and spacing.Poles average 550-foot spacing; cabinets connect each pole to the fiber backbone, while camera-lowering devices support maintenance and the custom six-camera clusters transmit video streams over a single gigabit Ethernet connection.Source: 2) Camera poles · p. 6

3) Video cameras:

The video system combines overlapping, synchronized cameras with centralized computing and processing software. It records timestamped video in manageable chunks and distributes tracking and trajectory reconstruction across dedicated hardware.

3) Video cameras:

  • Remote pan, tilt, and zoom alignment gives cameras overlapping 180-degree fields of view, while multiple cameras per pole extend coverage and reduce pole needs.Wider-field cameras were rejected after testing found edge distortion difficult to correct to the needed localization accuracy.Source: 3) Video cameras · p. 6

B. Network and Compute Hardware Architecture

  • A dedicated fiber connection delivers video to a data center where nine GPU-equipped nodes track vehicles across assigned cameras, and a postprocessing server assembles trajectories.The cluster also includes control, storage, development, and testing resources.Source: B. Network and Compute Hardware… · p. 7

C. Software Architecture

  • The software architecture has video-ingest, vehicle-detection-and-tracking, and trajectory-postprocessing modules managed by a control server.Configurations and metadata are registered before each run for record-keeping or reprocessing.Source: C. Software Architecture · p. 7

1) Video ingestion and recording:

  • Video is recorded in 10-minute MKV chunks with frame-exposure timestamps stored in presentation-timestamp metadata.At 1080p and 30 frames per second, all 276 camera streams use about 1 TB per recorded hour.Source: 1) Video ingestion and recording · p. 7

2) Vehicle Detection and Tracking:

I-24 MOTION detects vehicles in camera images, tracks detections in shared roadway coordinates, and reconciles fragments into trajectories. Its released data provide regularly sampled vehicle positions in a curvilinear coordinate system, with known accuracy qualifications.

2) Vehicle Detection and Tracking:

  • Crop-based Tracking uses predicted object-motion priors to process image crops on most frames, then matches detections across cameras using roadway-footprint intersection-over-union.The method uses a Kalman filter with linear dynamics and assumes constant velocity along the roadway and zero perpendicular velocity during initial tracking; multi-camera detection fusion avoids redundant tracking.Source: 2) Vehicle Detection and Tracking · pp. 7–8

3) Trajectory Post-processing:

  • Automatic postprocessing associates trajectory fragments and reconciles disturbed detections through derivative smoothing and outlier correction while minimally perturbing original detections.The pipeline currently does not yet resolve conflicts to ensure feasible inter-vehicle distances, which is planned for future development.Source: 3) Trajectory Post-processing · p. 8

IV. TRAJECTORY DATA / A. Data Description

  • Each dataset contains vehicle trajectories with attributes and motion information; positions describe the back-center footprint and are resampled at 25 Hz.Position data can be used to compute velocity, acceleration, and steering angle, and the overview includes known artifacts and preliminary accuracy analysis.Source: §IV · p. 8, A. Data Description · p. 9, Table III · p. 9

A. Data Description

  • The release contains ten weekday-morning datasets, typically four hours each, spanning crash- and debris-induced bottlenecks, travelling waves, and free flow.The November 25 dataset instead covers 11 hours; release information and statistics may change with future data updates.Source: A. Data Description · p. 9, Table IV · p. 9

B. Data Coordinate System

  • The curvilinear coordinate system aligns x with roadway direction and y with lateral position, making perpendicular roadway slices share a uniform x-coordinate.Speeds on the exterior of a curve may be underestimated by at most the roadway-width-to-curvature-radius ratio, typically under 5%; state-plane coordinates can be used for exact distances and speeds.Source: B. Data Coordinate System · pp. 9–10, Figure 7 · p. 10

C. Positional Accuracy

I-24 MOTION trajectories were assessed against manually labeled vehicle positions and onboard GPS, showing positional accuracy and lane-position information. The initial release also retains artifacts from occlusions, camera calibration shifts, packet loss, and tracking failures.

1) Manually Labeled Ground Truth:

2) GPS Data:

  • I-24 MOTION data shows lane-specific lateral-position peaks that onboard GPS does not show, indicating lane-position information GPS may require heavy filtering to obtain.The comparison used more than 600 GPS-equipped vehicle runs; the sensor’s reported circular error probable was 2.5 meters.Source: 2) GPS Data · pp. 10–11, Figure 9 · p. 12

D. Data Artifacts

  • Initial-release errors include missing-data bands from pole outages, overpass occlusion, camera homography shifts, and packet drops.Fragmented trajectories also arise from these artifacts and other tracking or post-processing failures; some errors remain regardless of algorithm, and manual correction does not scale to this dataset.Source: D. Data Artifacts · p. 12

E. Data Availability

I-24 MOTION releases processed trajectory data with metadata and expects to revise datasets as the instrument matures, while generally withholding raw video. Initial analyses illustrate traffic-wave observations from two weekdays and incident-related bottlenecks.

E. Data Availability

  • Processed daily trajectory files, metadata, and a README are released with DOI-based versioning, while raw video is generally not distributed because of privacy concerns and file size.The initial release corresponds to approximately 47 TB of video; datasets are released as-is and may be reprocessed as the instrument matures.Source: E. Data Availability · pp. 12–13

A. Traffic wave properties

  • The selected events included a severe rear-end crash followed by about 1.5 hours of upstream queuing, a 2.5-minute slowdown caused by a fallen object, and a sideswipe crash.The wave properties were estimated using visual inspection and techniques including wavelet transforms and cross-correlation.Source: A. Traffic wave properties · pp. 13–14
  • Traffic perturbations in the two weekday rush-hour datasets propagated upstream, generally at about 13 mph, while wave periods and speed ranges varied across events.Reported examples include Event A at 2.1 minutes and 14.8 mph, Event B at 5 minutes and 34 mph, and Event C at 1.8 minutes and 10.8 mph.Source: Table VII · p. 13, A. Traffic wave properties · p. 14, Figure 11 · p. 14
  • The released data can support speed analysis without long-distance extrapolation between fixed sensors, although data cleaning remains necessary.Examples also include small-scale bottlenecks, short-period waves, stopped vehicles, and off-ramp queues during otherwise free-flow conditions.Source: A. Traffic wave properties · p. 14, Figure 12 · p. 15

B. Fundamental diagrams (FDs)

I-24 MOTION’s spatial-temporal coverage supports examining traffic changes at finer scales and computing fundamental diagrams around incidents. The paper presents these analyses as initial illustrations, with data and instrument capabilities expected to improve.

B. Fundamental diagrams (FDs)

  • High-resolution spatial-temporal data supports finer-scale traffic analysis and fundamental diagrams at arbitrary locations around incidents.Figures 13 and 14 compare downstream and upstream incident conditions with same-location data from elsewhere that day.Source: B. Fundamental diagrams (FDs) · p. 15, Figure 13 · p. 15, Figure 14 · p. 16
  • The presented analyses are illustrative rather than comprehensive; the authors expect dataset improvements to support further investigations bridging microscopic and macroscopic scales.The incident diagrams use trajectory data and compare upstream and downstream speed, density, and flow conditions.Source: B. Fundamental diagrams (FDs) · p. 15

VI. CONCLUSION

  • The instrument’s processing algorithms were still far from final, and future work was planned to improve data quality and provide tools and methods for broader applications.The conclusion also describes the instrument as designed to support live traffic experiments, including automated-vehicle deployments and active traffic-management studies.Source: §VI · pp. 15–16

APPENDIX A I-24 MOTION INFRASTRUCTURE LOCATIONS

The appendices identify the camera-pole location table and example-trajectory tables, then define the coordinate systems and image-to-state-plane transformations used to represent vehicle data.

APPENDIX A I-24 MOTION INFRASTRUCTURE LOCATIONS

APPENDIX B EXAMPLE VEHICLE TRAJECTORY

  • The example trajectory is documented through a table of its first and last five points and a separate table of detailed trajectory information.Source: Table X · p. 18, Table IX · p. 18

APPENDIX C COORDINATE SYSTEM CONVERSION

  • Image coordinates use pixel row and column values; state-plane coordinates use Tennessee EPSG 2274 feet and a roadway-normal z-axis, while roadway coordinates follow a local longitudinal and lateral axis.Source: Appendix C · p. 19

A. Image to State Plane Conversion

  • Manually matched image pixels and surveyed state-plane landmarks fit a planar homography; a 3D perspective transform handles vehicle-box points above the ground plane.Source: A. Image to State Plane Conversion · p. 20

B. State Plane to Roadway Coordinate Conversion

State-plane vehicle boxes are converted to roadway coordinates by referencing a fitted roadway centerline and finding the closest longitudinal point on it.

  • A state-plane 3D bounding box is represented by eight corners; the roadway representation records position and vehicle length, width, and height.Source: B. State Plane to Roadway Coordinate… · p. 21
  • The box’s back-center point anchors the conversion: minimizing its distance to the centerline spline yields longitudinal coordinate x(r), and its distance from that point gives lateral coordinate y(r).A secondary spline provides an initial guess for the closest longitudinal coordinate; the optimization can be solved to arbitrary precision.Source: B. State Plane to Roadway Coordinate… · p. 21
  • The roadway centerline is a spline fitted midway between splines tracing the two interior yellow lines, with control points spaced at least 200 feet apart.Source: B. State Plane to Roadway Coordinate… · p. 21

C. Roadway to State Plane Coordinate Conversion

The inverse conversion constructs a state-plane vehicle box from roadway coordinates by locating its centerline position and applying lateral, longitudinal, and vertical offsets.

APPENDIX D ADDITIONAL SPACE TIME DIAGRAMS

The appendix provides additional time-space diagrams for freeway traffic on dates in November and December 2022.

  • The dated appendix labels also include diagrams for Nov 21–25, 2022, including Thanksgiving on Nov 24 and Black Friday on Nov 25.Source: Appendix D · p. 23
  • The listed weekdays run from Monday Nov 21 through Friday Nov 25, 2022.Source: Appendix D · p. 23
  • The additional westbound morning-rush-hour diagrams cover Nov 21, Nov 23, and Nov 29, followed by Dec 1 and Dec 2, 2022.Source: Figure 15 · p. 24

APPENDIX E LANE-DIS-AGGREGATED TIME-SPACE DIAGRAMS

The appendix presents lane-separated time-space diagrams for Wednesday, November 30, 2022, from 6:00 to 10:00 AM.

  • Fig. 16 presents lane-separated time-space diagrams for Wednesday, November 30, 2022, from 6:00 to 10:00 AM.Source: Figure 16 · p. 25

APPENDIX F TRAFFIC WAVE CALCULATIONS

The appendix estimates upstream wave propagation speed by cross-correlating speed signals and analyzes wave frequency over time with a continuous wavelet transform. The observed wave periods vary over the morning.

A. Wave propagation speed

  • For Nov. 22, 2022, the mean wave propagation speed was 12.8 mph (standard deviation 0.5 mph), estimated by cross-correlating speed signals at nearby locations.The method matches the first non-trivial peaks after shifting one signal relative to the other, then divides the time shift by the distance between locations.Source: A. Wave propagation speed · p. 26

B. Wave frequency analysis

  • The continuous wavelet transform convolves the speed time series with Morlet mother-wavelet functions, using continuously varying scale and translation factors.The resulting representation is overcomplete.Source: B. Wave frequency analysis · p. 26
  • The scaleogram shows nonstationary oscillations: wave periods are 6.7 min during 6:50–7:30 AM and near 9 min during 8:30–9:30 AM.Source: Figure 17 · p. 26, B. Wave frequency analysis · p. 27

Related papers

Loading 2301.11198v2…