I-24 MOTION Testbed for Traffic Science

• I-24 MOTION is an instrumented freeway testbed that generates comprehensive, high-fidelity vehicle trajectory data using a network of synchronized 4K cameras.
• It employs advanced calibration, data association, and sub-frame timing techniques to enable precise multi-scale analyses from individual vehicle dynamics to corridor-level traffic flow.
• The testbed supports multiple data products and simulation models, underpinning research in areas from microscopic vehicle tracking to macroscopic traffic and active management studies.

I-24 MOTION, short for Interstate-24 MObility Technology Interstate Observation Network, is a cyber-physical instrument and open-road, instrumented freeway testbed on Interstate 24 near Nashville, Tennessee, built to generate complete vehicle trajectory data at freeway scale from synchronized multi-camera video (Gloudemans et al., 2023). In the associated literature, it functions both as sensing infrastructure and as the basis for multiple public or planned data products, including corridor-scale trajectory releases, the Interstate-24-3D (I24-3D) benchmark for 3D multi-camera vehicle tracking, virtual trajectories derived from INCEPTION v1.0.0, and the I-24 MOTION Scenario Dataset (I24-MSD) for generative microscopic simulation (Gloudemans et al., 2023, Ji et al., 2023, Jayawardana et al., 10 Aug 2025).

1. Scientific purpose and corridor setting

The testbed was designed to enable traffic science across microscopic and macroscopic levels by producing complete vehicle trajectory data rather than the spatially sparse measurements typical of static loop-detector systems (Gloudemans et al., 2023). Its stated motivation is to observe how contemporary traffic dynamics, including the influence of connected and automated vehicles and active traffic management, manifest across time and space. A closely related motivation in the I24-3D benchmark paper is the need for high-fidelity, multi-camera 3D observations of traffic streams so that researchers can study both local vehicle-level and global traffic-level dynamics; that paper explicitly notes that controllers optimized solely for an ego-vehicle can induce rippling instabilities in traffic, whereas controllers that consider global traffic objectives can dissipate congestion waves (Gloudemans et al., 2023).

The instrumented corridor lies on Interstate 24 near Nashville, Tennessee, spanning approximately 4.2 miles (6.75 km) from Mill Creek (postmile 58.8) to milemarker 62.8 on a 4–5 lane each-direction freeway with frequently observed congestion (Gloudemans et al., 2023). The site has an annual average daily traffic of approximately 150,000 vehicles per day, with 10–15% heavy trucks, and approximately 230 million vehicle miles of travel occur within the I-24 MOTION segment annually. Typical morning congestion includes recurring oscillatory traffic with waves traveling upstream relative to traffic at approximately 12–13 mph. The corridor includes three overpasses, one underpass, and three interchanges with 13 entrance/exit ramps, and the testbed is co-located with TDOT’s I-24 SMART Corridor deployment, which includes variable speed limits, lane control, and ramp metering (Gloudemans et al., 2023).

This combination of freeway scale, recurring congestion, and corridor control infrastructure gives the testbed an explicitly systems-level orientation. The published use cases range from car-following and lane-changing analysis to wave propagation, incident-localized fundamental diagrams, active traffic management evaluation, and mixed-autonomy studies.

2. Physical instrumentation, networking, and synchronization

The instrument comprises 276 pole-mounted cameras distributed on 40 steel poles, with the roadway described in the virtual-trajectory paper as 8–10 lanes wide (Ji et al., 2023). In the original instrument paper, the poles are 110–135 ft tall, spaced at roughly 500–600 ft, with average spacing 550 ft, minimum 425 ft, and maximum 625 ft (Gloudemans et al., 2023). Each pole carries a custom camera cluster of six 4K-resolution pan/tilt/zoom IP cameras powered via PoE+, oriented orthogonal to the roadway. The PTZ capability is used to produce a 180-degree overlapping field of view across cameras on each pole and between poles, and six additional clusters are positioned at interchanges for ramps in future trajectory generation (Gloudemans et al., 2023).

Mechanical and electrical design are treated as part of the sensing problem. The poles and foundations were designed so that total deflection at the top is less than 1.5 inches in a 30 mph wind. A camera lowering device allows safe maintenance by winch and aggregates six 4K streams through a single gigabit Ethernet connection. Each pole cabinet includes a fiber patch panel, network switch, and two DC power supplies, while redundant Ethernet and power connections are accessible at ground level (Gloudemans et al., 2023).

The communications and compute path is similarly explicit. Every pole maintains a one gigabit per second link to an aggregation switch on the TDOT network, and the Vanderbilt University data center receives all feeds over a dedicated 40 Gbps fiber connection. The compute cluster includes a system control server, nine processing nodes each with eight GPUs for decoding and tracking, a post-processing server for trajectory assembly, and two storage arrays for trajectory data and logs. Incoming video is buffered on its respective node and discarded after processing (Gloudemans et al., 2023).

Time synchronization is central because multi-camera association depends on sub-frame timing consistency. Cameras are synchronized via network time protocol to primary and secondary stratum-1 GPS-based time servers at the TDOT hub building, resynchronizing roughly every 15 minutes. Camera firmware provides per-frame timestamps with approximately 10 microsecond accuracy relative to the camera clock, while exact inter-camera capture times can still differ. Typical frame-to-frame timing is 33.3 ms, with duplicated or skipped frames possible; the trajectory extraction pipeline compensates using accurate timestamps rather than assuming perfectly uniform frame intervals (Gloudemans et al., 2023).

3. Coordinate systems, calibration, and trajectory reconstruction

I-24 MOTION uses three spatial frames: Image Coordinates in pixels, State Plane Coordinates in EPSG 2274 Tennessee State Plane with the road surface at $z^{(st)}=0$, and Roadway Coordinates defined by a second-order spline along the median (Gloudemans et al., 2023). In the roadway frame, longitudinal motion follows the spline and lateral position is locally perpendicular to it; the frame is left-hand rule with positive $y$ on the eastbound side, and $x^{(r)}$ is offset so that postmile 60 equals $5280 \times 60 = 316800$ ft.

Image-to-ground calibration is based on planar correspondences between visible lane markings or landmarks in the image and surveyed points in state plane coordinates. The paper defines the perspective transform as

$$s \begin{bmatrix} x_i^{(st)}\ y_i^{(st)}\ 1 \end{bmatrix} \sim \mathcal{H} \begin{bmatrix} x_i^{(im)}\ y_i^{(im)}\ 1 \end{bmatrix},$$

with $\mathcal{H}$ estimated by minimizing reprojection error using OpenCV find_homography (Gloudemans et al., 2023). To project 3D vehicle prisms into the image, the paper introduces a $3 \times 4$ projection matrix $\mathcal{P}$, with its third column encoding the vanishing point for the $z$-axis. Roadway coordinates are then obtained by projecting the back-bottom center reference point of a vehicle prism to the closest point on the roadway spline:

$$x_o^{(r)} = \arg\min_{x^{(r)}} \left(\mathrm{dist}(F(x^{(r)}), \mathbf{o}_c^{(st)})\right), \qquad y_o^{(r)} = \min_{x^{(r)}} \left(\mathrm{dist}(F(x^{(r)}), \mathbf{o}_c^{(st)})\right).$$

Trajectory generation uses Crop-based Tracking, described by Gloudemans and Work, with RetinaNet and a ResNet-50 backbone as the base detector (Gloudemans et al., 2023). A Kalman filter with linear dynamics assumes constant velocity along roadway travel direction and zero velocity perpendicular to it during initial tracking. Affinity between detections and hypotheses is computed by intersection-over-union over vehicle footprints on the road plane rather than over image boxes, which enables multi-camera detection fusion given calibrated homographies. Post-processing then applies an online data association algorithm that solves a min-cost flow to match fragments belonging to the same object and a trajectory reconciliation quadratic program that smooths derivatives, corrects outliers, and enforces internal consistency among positions, speeds, and accelerations (Gloudemans et al., 2023).

The I24-3D benchmark adopts a simplified roadway-aligned storage convention specialized to annotation and evaluation. In that dataset, annotations are stored with longitudinal position $y$0, lateral position $y$1, a rear bottom-center anchor, and fixed vehicle dimensions in feet; lane width is 12 feet; and a flat planar road surface is assumed for per-camera homographies. The conventional forms referenced are the ground-plane homography $y$2 and the full perspective projection $y$3 (Gloudemans et al., 2023). The benchmark paper reports 1.24 ft average positional error between multi-camera annotations of the same vehicle and 0.5 ft average dimensional error.

4. Data products and release structure

The testbed has generated several distinct datasets and derived products, each emphasizing a different level of representation.

Release	Scope	Key characteristics
Initial public trajectory release (Gloudemans et al., 2023)	At least 4 hours for each of 10 days; 47 hr total	~600,000 vehicles; JSON trajectories + metadata; 25 Hz resampling
I24-3D (Gloudemans et al., 2023)	3 scenes; 16–17 cameras per scene; 57 minutes of video	877,000 3D boxes; 105,300 frames; 720 unique vehicles; VMT 209
INCEPTION v1.0.0 / virtual trajectories (Ji et al., 2023)	Typical morning commute on I-24 West, Nov 22, 2022, 6:00–9:00 AM	Raw trajectories at 25 Hz; 713 virtual trajectories per lane across four lanes; 2,852 total
I24-MSD (Jayawardana et al., 10 Aug 2025)	40 hours spanning 10 days	3.29 million tracks; 570k scenario segments of duration 9 s; 10 Hz; up to 32 vehicles per scenario

The initial public release distributes processed days as JSON collections of JSON-like trajectories with accompanying metadata, including scene homographies, algorithm settings, data attribute descriptions, and summary statistics (Gloudemans et al., 2023). Each vehicle trajectory includes a 12-byte BSON _id, class, timestamps, arrays of $y$4 and $y$5 positions in feet, starting and ending $y$6, length, width, height, direction, and a configuration ID linking to trajectory-generation settings. Trajectories record the 2D footprint of the back center of each vehicle and are resampled at 25 Hz to support exact timestamp-based indexing.

Class ontologies differ across releases. The corridor-scale trajectory datasets encode seven classes: sedan, midsize, pickup, van, semi, truck, and motorcycle (Gloudemans et al., 2023). I24-3D uses six classes: sedan, midsize, van, pickup, semi, and truck, and vehicles towing trailers are labeled by the towing vehicle’s class while the 3D box includes the trailer (Gloudemans et al., 2023). I24-MSD, by contrast, is summarized as a single-object-class vehicle dataset designed for simulation interoperability (Jayawardana et al., 10 Aug 2025).

Privacy and file-format decisions are release-specific. Raw video is generally not persisted or distributed because it may contain personally identifiable information and is too large for easy distribution; the initial release corresponds to approximately 47 TB of video files (Gloudemans et al., 2023). The I24-3D paper states that privacy considerations, file structure, example video links, and detailed formats are documented in appendices rather than fully reproduced in the main text (Gloudemans et al., 2023).

5. Validation and the I24-3D benchmark

Two distinct evaluation regimes appear in the literature. The first validates the corridor-scale trajectory extraction system itself; the second benchmarks generic multi-camera 3D tracking pipelines on a challenging annotated subset.

For corridor-scale validation, a manually labeled ground-truth dataset across 18 cameras was produced for free-flow and highly congested westbound scenarios, with over 600,000 individual vehicle positions labeled (Gloudemans et al., 2023). Predicted positions were matched to ground truth using IOU-based association. Reported results are MOTA = 0.93 in both scenarios, MOTP (IOU) = 0.73 in congestion and 0.72 in free-flow, Precision = 0.98 and Recall = 0.95 in congestion, and Precision = 0.97 and Recall = 0.96 in free-flow. Position error summaries report that 84% of predicted positions fall within 3 ft of ground truth and 36% within 1 ft; mean absolute errors are 1.7 ft longitudinally and 0.6 ft laterally, while dimension errors are less than 1.2 ft. Physical-feasibility checks also remain high, including feasible accelerations of 1.00 in both scenarios and feasible heading angles of 0.98 in congestion and 1.00 in free-flow (Gloudemans et al., 2023).

The I24-3D benchmark, by contrast, is explicitly constructed to stress multi-camera 3D object tracking under overlapping views, high speeds, and heavy occlusion (Gloudemans et al., 2023). It covers an approximately 2000-foot urban interstate segment near Nashville, recorded from 18 cameras mounted on three 110-foot roadside poles spaced at roughly 500 feet, though periodic outages leave 16–17 cameras per scene. The three scenes are: Scene 1, free-flow traffic, with 17 cameras, 90 seconds per stream, 45,900 annotated frames, 291,000 3D boxes, 324 unique trajectories, and VMT 118; Scene 2, slow traffic under snow conditions, with 16 cameras, 60 seconds, 30,600 frames, 146,000 boxes, 114 trajectories, and VMT 24.4; and Scene 3, congested traffic, with 16 cameras, 60 seconds, 28,800 frames, 440,000 boxes, 282 trajectories, and VMT 67.0.

The benchmark protocol splits each scene temporally into 80% training and 20% validation, trains detectors until convergence on RTX6000 GPUs, enforces synchronization at 1/60th second during evaluation, and compares spline-fit ground truth and predictions at unified discrete times nominally 30 Hz (Gloudemans et al., 2023). Metrics include CLEAR MOT, HOTA, Recall, Precision, MT, and ML, with a 30% 2D footprint IoU threshold on the road plane to accommodate residual synchronization error. Implemented components include the Single3D, Dual3D, and crop-based detectors; KIOU, ByteTracker, and crop-based tracking; and three cross-camera choices: Detection Fusion, Trajectory Fusion, and no fusion, plus the combined DF+TF variant.

The best overall pipeline is Dual3D + KIOU + Trajectory Fusion, which achieves HOTA 44.8%, MOTA 77.0%, Recall 83.0%, Precision 93.2%, GT% 91.7%, Pred% 92.3%, MT 63.8%, ML 8.8%, and ID switches per GT 0.52 (Gloudemans et al., 2023). Scene-wise, the same pipeline reaches HOTA 58.5 and MOTA 89.7 in free-flow, HOTA 46.9 and MOTA 77.7 in slow snow traffic, and HOTA 29.1 and MOTA 63.5 in congestion. The benchmark paper attributes the degradation in Scene 3 to long occlusions and lane-dependent visibility, with far lanes such as EB lane 1 and WB lane 4 showing more false negatives. It further notes that even with ground-truth single-camera tracklets, cross-camera rectification remains difficult in dense traffic, and in the best-performing pipeline roughly 5% of ground-truth objects are missed entirely.

Taken together, the two evaluations indicate different properties. The corridor-scale validation emphasizes the accuracy of the deployed extraction stack under the instrument’s operating assumptions, whereas I24-3D isolates the remaining difficulty of generic multi-camera association and trajectory continuity under stringent benchmark conditions.

6. Virtual trajectories, macroscopic fields, and lane-resolved analysis

A later line of work uses I-24 MOTION trajectory data as input to macroscopic reconstruction and lane-level analysis rather than direct microscopic track following (Ji et al., 2023). In that framework, raw lane-aligned 3D trajectories from INCEPTION v1.0.0 are processed into virtual trajectories by integrating a virtual vehicle through a smoothed speed field $y$7 rather than following noisy or fragmented empirical tracks.

The construction proceeds in three steps. First, a macroscopic speed field is computed using Edie’s definitions over shear-aligned spatiotemporal cells:

$y$8

Second, Adaptive Smoothing Method separates free-flow and congested components and smooths or interpolates along characteristic lines to fill gaps caused by occlusion or outages while preserving wave structure. Third, a virtual vehicle is integrated through the smoothed field according to

$y$9

The implementation uses forward Euler integration, with Fritsch–Carlson monotone piecewise cubic interpolation applied to $x^{(r)}$0 when the ODE time step is small relative to the grid (Ji et al., 2023).

The paper demonstrates this on I-24 Westbound for Tuesday, November 22, 2022, from 6:00–9:00 AM, using lane-by-lane grids with $x^{(r)}$1 miles and $x^{(r)}$2 seconds. A single 4-hour time-space diagram contains roughly 100–310 million points and about 70 GB uncompressed, so the authors recommend streaming Edie aggregation with preallocated TTT and TTD arrays rather than cellwise random access. For the 6:00–9:00 AM analysis window, 713 virtual trajectories were generated per lane across four lanes, for 2,852 total (Ji et al., 2023).

The demonstrated outputs are lane-level travel times and speed variability. In the cited morning window, the HOV lane had the shortest mean travel time, 7.86 min, but the highest mean speed standard deviation, 14.87 mph. Adjacent lanes had mean travel times of 8.26, 8.42, and 8.33 min with corresponding mean speed standard deviations of 13.39, 12.06, and 11.50 mph. Free-flow travel times cluster near approximately 4 minutes outside the congested period, while during congestion from roughly 6:20–9:00 AM travel times rise and fluctuate with departure time, peaking around approximately 7:45 AM. Departing only 10 minutes earlier or later can shift travel time by up to approximately 4 minutes because of different wave encounters (Ji et al., 2023).

The same paper defines stop-and-go waves as spatiotemporal patterns of alternating low and high speeds propagating upstream, with wave fronts of negative slope $x^{(r)}$3 and amplitude $x^{(r)}$4 characterized by the difference between local maximum and minimum speeds within a wave passage. Speed standard deviation is used as a proxy for fuel consumption, citing Barth and Boriboonsomsin, and the smooth $x^{(r)}$5 and continuous $x^{(r)}$6 profiles of virtual trajectories are positioned as suitable inputs for energy and emissions models (Ji et al., 2023).

7. Scenario modeling, persistent imperfections, and future directions

The I-24 MOTION testbed has also been repurposed for scenario-based generative microscopic traffic simulation through the I24-MSD dataset (Jayawardana et al., 10 Aug 2025). That release draws on 40 hours spanning 10 days on an instrumented westbound segment and summarizes 3.29 million tracks, average trajectory length 6.8 s, 570k scenario segments of duration 9 s, sample rate 10 Hz, single city, and single object class vehicles. Each scenario contains up to 32 vehicles over up to 9 seconds at 10 Hz and provides per-vehicle sequences of $x^{(r)}$7, $x^{(r)}$8, $x^{(r)}$9, and heading together with a vectorized road map aligned to the scene.

A central claim of that work is that infrastructure sensing yields a “messier” view of traffic than vehicle-mounted sensor suites, and that this imperfection should be treated as part of the learning problem rather than removed by aggressive sanitization (Jayawardana et al., 10 Aug 2025). The paper identifies calibration drift from thermal expansion or wind-induced pole tilt, transient occlusions from dust, debris, glare, or motion blur, frame drops and temporal discontinuities from bandwidth or firmware issues, multi-camera association failures causing ID switches and fragmentation, and map-trajectory misalignment as recurrent artifacts. To formalize this, it introduces noisy observations $5280 \times 60 = 316800$0 within a Hidden Markov Model framing of microscopic simulation and trains a GPT-style decoder-only Transformer, SMART, with robust losses including cross-entropy, label smoothing, focal loss, and symmetric cross-entropy.

On I24-MSD, SMART with noise-aware losses outperforms both IDM and Constant Speed baselines as well as standard cross-entropy training. The best reported configuration, CE + label smoothing, reaches Realism 0.7922, Kinematic 0.7406, Interactive 0.8300, Map-Based 0.7731, and minADE 1.3352, compared with SMART trained by cross-entropy alone at Realism 0.7698, Map-Based 0.7183, and minADE 2.0083 (Jayawardana et al., 10 Aug 2025). This suggests that the testbed’s value extends beyond trajectory measurement into standardized simulation inputs that preserve real-world sensing error.

The broader literature also documents persistent limitations in the underlying sensing system. These include missing pole data from network outages or pole damage, overpass occlusion, static homography errors caused by subtle pole deflections from temperature or sunlight, packet drops and frame corruptions, trajectory fragmentation, and calibration or synchronization imperfections (Gloudemans et al., 2023). In I24-3D, congestion and long occlusions remain especially problematic, and the benchmark explicitly recommends future methods that exploit multi-view geometry more explicitly in 3D, incorporate robust occlusion modeling and temporal reasoning, and handle lane-dependent visibility and speed (Gloudemans et al., 2023).

Published roadmaps remain expansive but technically specific. The instrument paper states that as the system matures, all trajectory data will be made publicly available at i24motion.org/data, and it emphasizes planned live experiments, including automated vehicle deployments to dampen jams, within the SMART Corridor context (Gloudemans et al., 2023). The I24-3D paper states that the authors plan to release a 3D multi-camera tracking challenge with new I-24 MOTION scenes and cameras (Gloudemans et al., 2023). The virtual-trajectory work points toward lane-wise wave characterization, energy and emissions modeling, lane management evaluation, and CAV control methods targeting wave mitigation (Ji et al., 2023). The scenario-modeling work encourages uncertainty-aware architectures, reinforcement-learning-based closed-loop fine-tuning, and broader robustness methods rather than heavier preprocessing (Jayawardana et al., 10 Aug 2025). Collectively, these directions position I-24 MOTION as both a measurement instrument and a continuously evolving benchmark environment for freeway traffic science.