OpenLKA
IEEE ITSC 2025 Journal extension under review · Accident Analysis & Prevention MIT-licensed code · data on request Up to 100 time-aligned signals · 100 Hz

OpenLKA

How lane keeping assist really performs on public roads

Lane keeping assist (LKA) is the feature that steers a car to stay centered in its lane. It ships on most new vehicles, and it works only as well as the road lets it. OpenLKA is the largest open record of how production LKA systems behave in real traffic: 389 hours of driving from 62 vehicle models, most of it recorded on Tampa Bay highways and rural roads.

See what the road does to LKA Paper (IEEE) Preprint GitHub Explore the signals Request the data

Yuhang Wang, Abdulaziz Alhuraish, Shengming Yuan, Hao Zhou · University of South Florida · MOTIF-Lab

The same sun-glare frame with the lane detector's lane lines, road edge and path drawn over it Forward camera frame on a Tampa interstate at sunset, with sun glare washing out the lane lines CameraLane detector
Sun glare on a Tampa interstate. Drag the handle: left is what the camera saw, right is what the lane model found. Lane-line confidence in this frame fell to 0.21 (left line) and 0.26 (right line) on a scale where 1.0 is certain.
389.1 h
of LKA driving
62
vehicle models
12
automakers
53
drivers
80+
cities

322 unique route logs · 30+ states and provinces · 83% recorded in Tampa Bay · decoded vehicle signals, synchronized video and scene labels · MIT license

01

Why this matters for state DoTs and road owners

Lane keeping assist reads painted lines with a camera. The road is half of the system.

It is already on the road

Lane keeping assist is on 86.3% of model-year-2023 vehicles, and continuous lane centering on 53.9% (PARTS, 2024). Drivers lean on it every day, whether or not the road was designed for it.

The road decides how well it works

Faded paint, white lines on pale concrete, glare, rain and tight curves all change what the car does. In our sample, four of the five most common conditions behind large drifts were about seeing the line.

Measured here, on Tampa Bay roads

Four lab drivers logged 323 hours on I-275, I-75 and rural two-lane roads from late 2023 through 2025, including the 2024 hurricane season. Community drivers added 66 hours from 30+ states and provinces.

A comma 3X recorder mounted behind the windshield of a Hyundai Ioniq 5, next to cockpit photos of six test cars: Kia EV6, Tesla Model 3, Honda Accord, Hyundai Ioniq 5, Ford Mustang Mach-E and Volkswagen Tiguan

How we measure it, in four steps

  1. 1
    Record. A dash-cam-sized recorder sits behind the mirror of an ordinary rental car. It logs the car's internal data network (the CAN bus) and the forward camera at the same time.
  2. 2
    Decode. We decode the hidden signals: whether lane keeping is engaged, steering angle and torque, speed, pedals. Every decoded signal is checked against the video.
  3. 3
    Measure. A camera-based lane model gives the car's position in the lane. We checked it against LiDAR ground truth: typical error 0.135 m.
  4. 4
    Tag. An AI model that reads images labels every scene in ten categories (rain, glare, faded paint, construction and more). We checked 2,000+ of its labels by hand.

Why it matters for your DoT. The same recorder and method can audit any corridor, in Florida or elsewhere, for how production lane keeping behaves there, before and after a marking or paving project.

02

Ten road conditions, one dataset

The same Tampa roads were driven again and again in different weather and light. Community drivers added snow, mountains and city traffic.

Interstate in heavy rain; spray from the vehicles ahead hides both lane lines
Heavy rain, Tampa
Snow-covered highway at dusk with the lane lines buried
Snow, community drive in Canada
Road edge lost on a wet surface (loop)
Rain: lines lost, system hands back control

Spray hides both lane lines. In our recordings LKA often disengaged in heavy rain, which means the driver gets the car back with no warning. Florida's afternoon storms make this a routine event, not an edge case.

Why it matters for your DoT. Every one of these conditions exists on a state highway system today. The dataset lets an engineer see how a production car reacts to each one, without a test track.

03

What the car sees

Pick a road condition, then drag the handle. Left is the camera frame, right is what the lane model found.

Sun glare frame with the lane detection overlay Sun glare on a Tampa interstate CameraLane detector

The lane model is the open-source openpilot lane detector, used here only as a camera-based measuring tool. Its lane-position estimate was checked against LiDAR ground truth on the OpenLane benchmark: typical error 0.135 m for lane position and 0.281 m for line distance. The drifts discussed on this page are two to nine times larger than that error.

04

The Tampa test fleet

Ordinary rental cars with factory driver assistance, recorded by the lab's own devices. Click a car to watch a 12-second clip around the moment the driver takes over.

Clips come from ADAS-TO, the takeover subset of the same recordings, trimmed to 12 seconds with the driver's takeover at the midpoint. The lab's recorder runs openpilot; in most clips its lane centering was engaged before the takeover, and in one Camry clip the vehicle's own system was. Each clip's caption states which. Hours per model are from the dataset paper where published. Source video is 526 × 330 pixels.

Why it matters for your DoT. These are cars people rent and buy today. The dataset shows how each one behaves on the same roads, under the same conditions.

05 · CAN lab

Inside the CAN bus: the car's own signals, time-aligned at 100 Hz

The recorder logs the car's CAN bus and the forward camera together. We turn that log into a 100 Hz table in which every row holds the car's own lane-keeping state, steering and speed next to what the camera-based lane model saw at the same instant.

83–100
time-aligned columns per decoded minute
100 Hz
CAN rate, resampled onto the speed clock
9
vehicle decoders, 7 verified DBC files
≈5 s
to decode one minute of driving
  1. 1
    Record. A comma 3X behind the windshield writes every CAN frame at 100 Hz and GPS at 1 Hz into a route log (rlog), and the forward camera at 526 × 330 into a synchronized camera file (qcamera). One minute is about 47 MB of log. For the stock clips below it ran as a read-only dash cam; on other drives it was also steering (openpilot), and those clips say so.
  2. 2
    Decode. openpilot's LogReader opens the segment. Raw CAN frames are decoded with a per-make DBC file; we release verified DBCs for seven makes. Out come the car's own signals: lane keeping active or faulted, cruise state, steering angle, driver and motor torque, pedals and, where the car broadcasts it, its lane-line detection.
  3. 3
    Align. Everything is resampled onto the 100 Hz speed clock (vEgo): states nearest-sampled, continuous values interpolated. The lane model's 33-point lane lines and confidences join the same rows. One minute becomes 6,000 rows by 83 to 100 columns, in about five seconds.

Signal Explorer: one clip, every decoded channel

Pick a drive; each tab names its recorder. The video, the car's own lane-keeping state, steering, the camera lane model and the raw CAN frames all run on one clock. Click anywhere on a chart to seek. A clip is labelled "stock" only when the recorder transmitted no steering frames at all; the one clip where openpilot was steering says so.

0.0°steering angle (CAN)
    GPS track at 1 Hz · bright = driven so far
    Recorder
    —
    Vehicle
    —
    Route · start
    —
    Assist state
    —
    Decoder · DBC
    —

    Raw CAN frames at the key moment (table)
    t (s)busaddressmessagebytesdecoded
    Sixty seconds of decoded data in heavy rain next to the camera frame: lane deviation, the car's own line-detection code, lane-model confidence and LKA active or disengaged shading
    Sixty seconds in heavy rain, from the paper (a Hyundai Ioniq 5). Right: the camera frame. Left: lane deviation (blue), the car's own line code (labelled 1 solid, 2 faded, 3 departure in the paper; green and orange), lane-model confidence (red, magenta) and LKA active (light) or disengaged (grey) shading.

    Three recorders, one method

    Three of the lab's own comma 3X recorders produced the clips on this page. Each has a fixed ID that appears in every file name; those IDs are the provenance key for every clip and table here.

    530075d26cad58e4

    Mach-E · Model 3 · Ioniq 5 · Niro EV · Maverick

    Hours
    52.1
    km
    2,516
    Route logs
    113
    bdda168c0c35fad7

    Main lab recorder: Tiguan · RAV4 · Camry · Accord Hybrid · Ioniq 5 · Model 3 · Model X · Niro · EV6 · Civic

    Hours
    122.6
    km
    6,766
    Route logs
    238
    d5a6fb2f1b849a62

    Kia EV6

    Hours
    5.6
    km
    268
    Route logs
    7

    These three devices hold 180 h, 9,549 km and 358 route logs on the lab server, counted per recording (hours and distance from the route metadata, routes from the dataset directory). The paper's Tampa figures (323 h, 127 unique routes) cover the whole campaign and de-duplicate repeated drives of the same road.

    What each make broadcasts, and what we recover

    Make
    Volkswagen✓✓✓✓✓✓✓✓
    Ford✓✓✓✗✓✓✓◐
    Toyota✓✓✓✓✗✓◐✓
    Honda✓✓✓✓✓✓✓✓
    Tesla◐✓◐◐✓✓✓◐
    Kia✓✓✓✗✓✓✗✓
    Hyundai✓✓✓✗✓✓✗✓

    ✓ broadcast on the bus and decoded   ◐ recovered through DBC integration and validation   ✗ not available. Source: the journal revision's signal-availability table. "Steer angle & torque" means at least one steering signal is available; pedals are normalized percentages. Hyundai's camera ECU (0x2A4) also broadcasts a 0–3 line-detection level per side on the Ioniq 5 and Niro, decoded after that table was drawn.

    Figure from the conference paper: per-make matrix of ten decoded signal families, with a CAN logger icon and the stock measurements it exposes: kinematics, steering, torque, lane-line detection and LKA state
    The matrix as drawn for the ITSC 2025 conference paper, with ten families (including clock and lane offset). The table above follows the journal revision, which regrouped the families and re-checked several cells.

    Nine vehicle decoders, seven verified DBC files

    Hyundai · Kia

    ioniq · kia_ev6 · niro

    hyundai_canfd.dbc

    • MDPS 0x0EA → LKA_ACTIVE, LKA_FAULT, column torque
    • LKAS 0x050 / LFA 0x12A → LKA_ICON, TORQUE_REQUEST
    • STEERING_SENSORS 0x125 → angle, rate
    • CAM 0x2A4 → left / right line-detection level (Ioniq 5, Niro; not broadcast on the EV6)
    Toyota

    toyota

    final_toyota_canfd.dbc

    • LKAS_HUD 0x412 → LKAS_STATUS, LEFT / RIGHT_LINE, LDA_ALERT
    • LTA_RELATED 0x371 → LTA_STEER_REQUEST
    • STEER_TORQUE_SENSOR 0x260 → driver and EPS torque
    Ford

    mache

    ford_canfd.dbc

    • LateralMotionControl2 0x3D6 → LatCtl_D2_Rq, path offset
    • Lane_Assist_Data1 0x3CA → LkaActvStats_D2_Req (lane-departure aid)
    • Battery, regen and radar distance
    Tesla

    tesla_model3

    tesla-model-3.dbc

    • DAS_status 0x399 → DAS_autopilotState, ACC state
    • DIR_torque 0x108 → motor torque, actual and requested
    • Steering angle, lateral acceleration, pedals, battery and iBooster brake
    • Autopilot was not engaged in any lab Model 3 recording, so no Tesla clip appears above.
    Honda

    accord · civic

    final_cleaned_extended_honda_accord_acc_lka.dbc · honda_civic_ex_2022_can.dbc

    • STEER_STATUS 0x18F → STEER_CONTROL_ACTIVE
    • LKAS_HUD 0x33D → LKAS_OFF, LANE_LINES, SOLID / DASHED
    • Lane lines 0x240–0x244 → offset, probability, visible distance
    Volkswagen

    volkswagen

    vw_mqb_2010.dbc

    • LH_EPS_03 0x09F → EPS_HCA_Status, driver torque
    • HCA_01 0x126 → status, LM offset
    • LDW_02 0x397 → distance and time to line
    Three-step DBC pipeline: candidate DBC collection, a signal-parsing loop, then empirical cross-checks against video and physics; pedal position against acceleration gives R = 0.746 and steering angle against curvature gives R = −1.000
    Decoded, not guessed. Each DBC is checked three ways: bit-level parsing against public signal definitions, visual checks against the video, and cross-signal correlation. Accelerator pedal against measured acceleration: R = 0.746. Steering angle against the lane model's path curvature: R = −1.000 (opposite sign convention). The DBC files and checks are on GitHub.
    Glossary of the decoded columns used on this page
    vEgo · aEgo
    Speed (m/s) and acceleration (m/s²) from the car. vEgo is the 100 Hz clock every other column is aligned to.
    lka_active · lka_status · lta_steer_request · LatCtl_D2_Rq · STEER_CONTROL_ACTIVE · EPS_HCA_Status · das_autopilot_state
    Is the car's own lane centering steering right now? Hyundai/Kia (MDPS 0x0EA), Toyota, Ford, Honda, Volkswagen (EPS_HCA_Status 5 = active) and Tesla each broadcast it differently; the Explorer shows the make's own signal as "LKA".
    acc_enable · op_enable · op_tx
    The car's adaptive-cruise state; whether the recorder's openpilot was engaged; whether it was transmitting steering commands (0 on dash-cam-mode drives). Clips labelled "stock" have op_tx = 0 throughout.
    left_lane_line · right_lane_line
    The car's own line code where the make broadcasts it. Toyota (0x412): 0 none, 1 solid, 2 faded, 3 orange (departure warning). Hyundai/Kia (0x2A4, Ioniq 5 and Niro): 0 not detected, then a 1–3 detection level that fell from 3 to 1 as the right marking faded in our data; the paper labels the same code 1 solid, 2 faded, 3 departure.
    op_lane_left_prob · op_lane_right_prob
    Camera lane-model confidence, 0 to 1. These are the bars in "What the car sees".
    steering_angle · steering_rate
    Steering-wheel angle (degrees, positive left) and rate from CAN.
    mdps_steering_torque · steer_torque_driver · EPS_Lenkmoment
    Driver torque on the column (Hyundai/Kia, Toyota, Volkswagen). A takeover shows as a spike. TORQUE_REQUEST, STEER_TORQUE_CMD and HCA_01_LM_Offset are the assist's commanded torque.
    op_left_laneline · op_right_laneline
    33-point lane-line positions (m) from the lane model. Lane deviation is the mean of the two lines at the car: positive means the car sits left of the lane center.
    easternTime
    Wall-clock time of the row (US Eastern).

    Why it matters for your DoT. An engineer can read the car's own lane-keeping state on a corridor, second by second, with no instrumentation beyond a windshield recorder, and compare it with what the road looked like at that instant.

    06

    Where lane keeping struggles

    On good roads the systems are steadier than people. When the line is hard to see, drifts get large.

    0.107 vs 0.140 m

    Spread of lane position under normal conditions, LKA versus human drivers. The systems are about 24% steadier when the road is clear.

    0.3 m

    Typical worst drift on a route (the 95th-percentile deviation, median across routes: 0.32 m left, 0.28 m right).

    > 0.8 m

    Some drifts exceed 0.8 m (2.6 ft), seen across automakers. That puts a tire on the line.

    > 0.2 m/s

    Sideways drift rates above 0.2 m/s are common. At that rate a car uses up half its lane margin in about two seconds.

    What was going on when the car drifted a lot

    Share of large-deviation samples tagged with each condition. A sample can carry more than one tag.

    • Lane line hidden by traffic19.7%
    • Low line contrast14.6%
    • Low light12.1%
    • Faded markings11.5%
    • Sharp curve9.6%
    • High speed6.4%
    • Construction5.7%
    • Sun glare5.5%
    • Road surface5.2%

    Seeing the line   Road geometry   Other

    Under the four line-visibility conditions, the typical worst drift per route rises above 0.5 m (median 95th percentile 0.52 to 0.56 m).

    How much room is there in a lane?

    A 12-foot lane and a typical car, seen from above. Markers show the drifts measured in the dataset.

    Top-down lane diagram: a 3.66 m lane with a 1.9 m car. A scale below the lane marks how far the car's center has drifted: 0.25 m noticeable, 0.5 m typical worst drift under poor markings, 0.65 m critical, 0.88 m tire on the line, 1.2 m sharp-curve case. 0 0.25 0.5 0.65 0.88 1.2 m drift of the car's center from the lane center (m) 3.66 m (12 ft) lane · 1.9 m car · about 0.9 m of margin to each line

    0.25 m noticeable   0.5 m typical worst drift under poor markings   0.65 m critical   0.88 m tire on the line   1.2 m sharp-curve case

    About 0.9 m (3 ft) separates a centered car from the line. Under poor markings the typical worst drift uses more than half of it. The worst cases in the dataset put a tire on the line.

    Four failure patterns, one case each

    Camera frames and lane-deviation trace: white lines on pale concrete beside a truck; lane-line detection drops and the car drifts about 0.8 m toward the truck
    Perception · 0.8 m beside a truck

    The camera could not separate white paint from white concrete

    With a truck alongside, the detector's confidence dropped and the car drifted about 0.8 m (2.6 ft) toward the truck.

    Road lever: contrast between paint and pavement.

    Two drives at lane transitions: a Hyundai Ioniq 5 loses the line at a merge, drifts and disengages, while a Ford Mustang Mach-E stays stable at a diverge
    Planning · merge and diverge

    Where lines split, the car has to guess which line is its lane

    One Ioniq 5 lost the right line at a merge, drifted and switched itself off. A Mustang Mach-E held the lane at a diverge. One event each, not a ranking of brands.

    Road lever: continuous, unambiguous markings through merges and diverges.

    A Hyundai Ioniq 5 on a sharp curve: lane deviation reaches about 1.2 m while lateral acceleration saturates below what the curve demands
    Control · 1.2 m on a sharp curve

    The camera saw the line; the steering could not follow it

    On a sharp two-lane curve the car drifted 1.2 m (3.9 ft) toward the outside while both lines stayed detected. The system's lateral authority ran out before the curve did.

    Road lever: curve signing, advisory speeds and geometry that production systems can follow.

    Three drives where conditions stack: a sharp curve with worn paint, a lane transition with worn paint, and heavy rain at a transition; drifts of about 1.2, 0.8 and 0.7 m
    Compounding · 0.7 to 1.2 m

    When conditions stack, drifts grow

    Worn paint on a curve, worn paint at a transition, rain at a transition: three drives, drifts of about 1.2, 0.8 and 0.7 m. Two ordinary problems in the same place behave like one big one.

    Road lever: fix markings first where geometry is already demanding.

    Sharper curve, bigger outward drift

    In the sharp-curve runs we analysed, drift grew roughly in step with curvature. Curvature is 1 divided by the curve radius. Move the slider.

    Curvature0.010 1/m
    Curvature-related drift0.08 m (0.3 ft)
    Bandwithin the normal band

    Fit from the dataset's curve segments: lane deviation ≈ −8.33 × curvature + 0.21 m (R² = 0.67, so curvature explains about two thirds of the variation). The slider shows the curvature term only, and stays inside the fitted range (radius 10 m and up). The sign only tells the direction: toward the outside of the curve. Other roads and speeds will differ.

    Scatter plot of lane deviation against curvature with a fitted line of slope minus 8.327 and R squared 0.673
    Measured drift versus curvature across the sharp-curve runs. Each color is one vehicle; the line is the fit.
    Violin plots of lane position for human drivers and eight automakers under normal conditions; the automated systems show a tighter spread than human drivers
    Normal conditions, human drivers versus eight automakers (1,087 samples). The systems hold a tighter line than people when the road is clear: spread 0.107 m versus 0.140 m.

    Why it matters for your DoT. Three of the four failure patterns have a road-side lever. The dataset tells you which lever applies on which kind of road.

    07

    What the data suggests for road owners

    Three levers an agency controls, each tied to what we measured. These are associations observed in our sample, not controlled experiments.

    01

    Raise line-to-pavement contrast

    White paint on light concrete and faded yellow lines were behind the lowest detector confidence and some of the largest drifts in our sample. Contrast is the cheapest lever, and it helps human drivers too.

    See the low-contrast frame
    02

    Make merges and diverges unambiguous

    Where lines split, the car has to pick a line. Continuous edge lines and early lane assignment through transitions give the camera one clear answer.

    See the merge case
    03

    Measure before and after

    A dash-cam-sized recorder plus decoded vehicle signals shows how production lane keeping behaves on a specific road section. Run it before a project and again after, on any corridor.

    See the method

    Why it matters for your DoT. Lane keeping is becoming standard equipment. Roads that production systems can read are a maintenance target you can set and check.

    08

    The dataset: scale and diversity

    Large enough to compare automakers, dense enough on one region's roads (Tampa Bay) to study them corridor by corridor.

    Three maps of the Tampa test routes: I-275 through Tampa, a rural two-lane loop in Pasco County, and I-75 to Temple Terrace
    The Tampa campaign: I-275, a rural two-lane loop and I-75, driven repeatedly from late 2023 through 2025 by four lab drivers. 323.0 hours, 127 unique routes.
    World map of recording locations and a bar chart of hours by region: United States 349.1 hours, then Europe, Canada, Taiwan and other regions
    Community drives add 66.1 hours from 49 drivers across 30+ states and provinces, including Canada, Europe and Taiwan.

    Hours by region

    United States 349.1 h (89.7%)   Europe 11.5 h   Canada 10.4 h   Taiwan 6.7 h   Other 11.4 h

    Hours by source

    Tampa lab tests 323.0 h, 4 drivers, 127 routes   Community 66.1 h, 49 drivers, 195 routes

    Hours by vehicle model

    • Hyundai Ioniq 5 (two configurations)89.7 h
    • Kia Niro EV48.3 h
    • Tesla Model 347.8 h
    • Kia EV6 (two entries)51.3 h
    • Toyota Camry (two configurations)41.2 h
    • Ford Mustang Mach-E30.3 h
    • Honda Accord Hybrid19.0 h
    • Other models61.6 h

    What is in a recording

    VideoForward camera at 526 × 330 for every drive; 1928 × 1208 HD and a wide fisheye view for the lab's own drives. Faces and plates blurred.
    Vehicle signalsCAN bus logs at 10 and 100 Hz, decoded: lane keeping and cruise status, steering angle and torque, speed, acceleration, pedals.
    Lane geometryLeft and right line distance, road edge and curvature from the on-device lane model, checked against LiDAR ground truth.
    Scene labelsTen categories (weather, light, marking quality, construction, curves and more) from a vision-language model, checked against 2,000+ hand labels.

    How it compares

    Hours of driving in public datasets. OpenLKA is the only one that pairs video with decoded vehicle signals and scene labels for lane keeping.

    DatasetHoursLKA focus
    OpenLKA389video + vehicle signals + scene labels
    nuScenes, BDD100Kabout 1,200no
    Waymo Open570+no
    HDD104no
    BDD-X77no
    comma2k1933no
    KITTI1.5no

    Subsets

    • Main 389.1 h, 62 models, all modalities (on request, see Access)
    • Normal clear-road driving, the human-versus-LKA baseline (350 samples, 2.9 h)
    • Failure large deviations and disengagements with scene labels
    • Alert the LKAlert subset for warning research
    Open samples (Dropbox) Request the full dataset

    Why it matters for your DoT. 83% of the hours were recorded on Florida roads, so the statistics describe one state's markings, weather and traffic in depth. Community drives from 30+ states and provinces show how the same systems behave elsewhere.

    09

    Our work and outputs

    One Tampa campaign grew into an open dataset, peer-reviewed papers and a family of six sibling datasets built on the same recording method.

    389.1 h
    decoded and time-aligned
    7
    automakers with verified signal decoding
    2000+
    hand-checked scene labels
    6
    sibling datasets
    Four scenes comparing the camera-based lane lines with LiDAR ground truth: a curve, an occluded line, night and a diverge
    Checked against LiDAR. Lane position error 0.135 m, line-distance error 0.281 m on the OpenLane benchmark.
    Four-step pipeline for decoding vehicle signals: public signal definitions, bit-level parsing, checks against video and other signals, verified definitions
    Decoded, not guessed. Every vehicle signal is consolidated from public definitions and checked against the video and other signals.
    Annotation pipeline: three consecutive frames go to a vision-language model with a structured prompt and return scene labels
    Labeled at scale, checked by people. Ten scene categories from a vision-language model, 2,000+ labels verified by hand.
    1. Late 2023
      Tampa road tests begin. Lab recorders ride in rental cars on I-275, I-75 and rural roads.
    2. 2024
      The fleet grows to the models above. Repeated passes in all weather, through the 2024 hurricane season.
    3. May 2025
      Preprint. arXiv 2505.09092 describes the dataset and first findings.
    4. Jun 2025
      Accepted at IEEE ITSC 2025.
    5. Jul 2025
      From OpenLKA to LKAlert. Abstract accepted at INFORMS 2025, Atlanta.
    6. 2025 to 2026
      Journal extension under review at Accident Analysis & Prevention: human-driving baseline, cross-automaker benchmark, LiDAR validation.
    7. 2026
      A dataset family on the same pipeline. ADAS-TO, BATON, DriveDNA, DriveMotion, VLAlert and TriDrive.

    The dataset family

    IEEE ITSC 2026

    ADAS-TO

    16,446 takeover-centered clips from 364 drivers. The fleet clips on this page come from it.
    Preprint 2026

    BATON

    136.6 h of bidirectional human and automation control transitions with video, CAN, radar and GPS.
    Preprint 2026

    DriveDNA

    Driving-style identification: 4,121 drives from 465 drivers, 975 h.
    Dataset 2026

    DriveMotion

    400 h of privacy-reduced in-cabin body motion for driver-motion forecasting.
    CoRL 2026

    VLAlert

    Observe before you alert: a vision-language framework for adaptive driver alerting, with a 192,892-tick benchmark.
    arXiv 2026

    TriDrive

    A world model that forecasts driver, vehicle and road demands in real time on a dash-mounted device.

    Why it matters for your DoT. The team has run this recording method for two years on Tampa Bay roads. The pipeline, the people and the fleet are in place, and the method travels: a recorder, a rental car and a corridor.

    10 · Access

    Access the data

    Samples are open now. The full decoded dataset is shared on request, so we know who is using it and can point you to the right subset.

    Diagram: daily driving feeds logging, reverse engineering and VLM annotation, which feed an online database shared through Dropbox and GitHub and the OpenLKA site
    From a daily drive to a download: logging, CAN decoding, VLM labeling, release on Dropbox and GitHub.
    Available now

    Open samples

    Normal, Failure and Alert pre-release folders on Dropbox; code, DBC files and download.sh --partial on GitHub.

    Normal Failure Alert GitHub
    On request

    Full decoded dataset

    Decoded per-route CSVs (100 Hz where the full route log exists, 10 Hz otherwise) with synchronized 526 × 330 video for all 389 h. Request, review, then a Google Drive or Dropbox link by email.

    Request access
    Research agreement

    Raw logs

    Raw rlog and qcamera files, plus HD and fisheye video for the lab fleet. Requires a signed research agreement between institutions.

    Start a conversation

    Request by email

    We review requests within 5 business days and reply from yuhangw@usf.edu with a download link or questions. Tell us who you are, your affiliation, what you intend to do with the data and which subset you need.

    Email a request

    Request form: coming soon, use email. Last release: pre-release folders (Normal / Failure / Alert) on Dropbox.

    Terms of use

    • Code and DBC files: MIT license.
    • Data: cite the ITSC 2025 paper and, once published, the Accident Analysis & Prevention extension; attribute OpenLKA in derived work.
    • Do not attempt to re-identify drivers, vehicles or people in the video.
    • Do not redistribute raw video. Share derived tables only with these terms attached.
    What is in a full-dataset CSV?

    One row per 0.01 s (0.1 s for routes that only have the 10 Hz log): the car's own lane-keeping and cruise state, steering angle and torque, speed, acceleration, pedals, the lane model's lane-line positions and confidences, GPS and wall-clock time. See the glossary in the CAN lab section.

    Can I get just one make, one recorder or one corridor?

    Yes. Name it in your request; we can cut by make, recorder ID, date or route.

    Why a review step?

    To know who uses the data, to keep raw video under the terms above, and to point you to the subset that fits your question.

    Will there be a DOI?

    A versioned snapshot will be archived with a DOI alongside the journal publication.

    Why it matters for your DoT. An engineer can start with the open samples today and request the decoded routes for the corridors they manage.

    11

    Read the paper, cite the data

    Code and DBC files are MIT-licensed; the data is shared under the OpenLKA terms above. For state DoTs and road owners: we can run the same method on your corridors.

    GitHub Project page IEEE ITSC 2025 2505.09092

    BibTeX

    @inproceedings{wang2025openlka,
  title     = {OpenLKA: An Open Dataset of Lane Keeping Assist from Production Vehicles Under Real-World Driving Conditions},
  author    = {Wang, Yuhang and Alhuraish, Abdulaziz and Yuan, Shengming and Zhou, Hao},
  booktitle = {IEEE International Conference on Intelligent Transportation Systems (ITSC)},
  year      = {2025}
}