You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

VideoTraj-Long v0.1

Long, trimmed forward-camera video of naturalistic drives, time-aligned to the CANTraj GPS tracks, for video-only trajectory recovery over whole trips

GitHub HF Dataset HF Dataset

3,689 drives by 440 drivers, 93,942 one-minute video segments, 1,566 hours of 20 fps road video (112,728,680 frames, 182.1 GB), each drive paired with its GPS track on a shared time axis. Every drive keeps the pseudonym it has in DriveDNA and CANTraj.

Task

Given the forward road video of a whole drive, its start position and (optionally) a vector road map, recover the vehicle trajectory: causally while the video plays, with revisions, and offline after the drive. The reference is the GPS track of the same drive on the same clock. The evaluation protocol (information levels P0/P1, output classes A/B/C, failure-aware Acc@100m, whole-drive and long-horizon cohorts, splits) lives in protocol.md of the GitHub repository. This repository holds the data only.

TL;DR

  • What: for each released drive, the kept one-minute qcamera segments as videos/driver_XXX/drive_YYY/segN.mp4 (H.264, 526x330 at 20 fps; 480x360 for the 7 comma two drives), one GPS CSV gps/driver_XXX/drive_YYY.csv, and manifests with per-drive and per-segment metadata, coarse region labels, camera groups and splits.
  • Time axis: time_s = 0 is the first frame of the first kept segment. Frame i of segment N is at t_start_s(N) + pts_i, where pts_i is the frame's presentation time inside the mp4 (the container timeline restarts at 0 for each segment, 50 ms nominal step). t_offset_cantraj_s converts to the CANTraj and DriveDNA axis: time_s_cantraj = time_s + t_offset_cantraj_s.
  • Trimmed: the first and last two minutes of every drive are removed on the video's own clock as well as on the CANTraj clock (details below). Holes inside a drive are kept as holes and listed.
  • No absolute time, no audio, no cabin camera, no CAN: the month of recording (year_month) is the only calendar information; the CAN, IMU and 10 Hz vehicle tables of the same drives are in DriveDNA and CANTraj under the same pseudonyms.
  • Scope: 3,689 of the 3,940 CANTraj main drives (the 251 others have no video file inside their trimmed window, mostly short drives). The 48 Controlled-supplement drives are not included.
  • Access: gated, research use only, under the DriveDNA research license; approval is manual.

What is released

kind files count size
video segments videos/driver_XXX/drive_YYY/segN.mp4 93,942 112,728,680 frames, 1,566 h 182.1 GB
GPS tracks gps/driver_XXX/drive_YYY.csv 3,689 11,315,494 rows 1.2 GB
manifests/ (drives, segments, splits, camera groups, region counts, checksums) 12 3,940 drive rows, 93,983 segment rows small

97,647 files in total. Median drive: 15 minutes of video; 331 drives are at least one hour long, 97 at least two hours, 14 at least four hours and 5 at least six hours.

Not included: the driver-facing camera, the 1928x1208 wide road camera, audio (removed from the 11,255 source segments that carried a cabin microphone track), CAN or IMU signals, absolute timestamps, and anything from the Controlled supplement.

Relation to DriveDNA and CANTraj

  • driver_XXX / drive_YYY are the pseudonyms of DriveDNA (10 Hz vehicle signals, untrimmed first video segments) and CANTraj (trimmed GPS tracks and raw CAN/IMU event tables). The same drive has the same name everywhere; there is no other link and no identifier of the vehicle, device, route or person in any file.
  • The GPS file of a drive is the public CANTraj GPS CSV of that drive, cut to the released video span and shifted by t_offset_cantraj_s; every other field is byte for byte the CANTraj value. The CANTraj localizer and GNSS files are not duplicated here.
  • The split columns (split_main, split_cross_metro, route_family, containment statistics) are copied from the CANTraj benchmark manifests, so results are comparable drive by drive with the CAN-to-GPS baselines.
  • DriveDNA already publishes the untrimmed first video segment of its drives under the same gate; this repository adds alignment and length, not new exposure (see Privacy).

Trimming and alignment guarantees

For every released drive the following hold and are checked by release/verify_release.py (report in the code repository):

  1. t_offset_cantraj_s >= 120: the released video starts at least two minutes after the zero of the CANTraj table (the first vehicle-state message of the anchor segment).
  2. cantraj_duration_s - (t_offset_cantraj_s + duration_video_s) >= 120: it ends at least two minutes before the end of the table.
  3. The head removed on the video's own clock is at least two minutes as well (trim_head_s; within 2 s of 120 s for drives whose camera started up to two seconds after the table zero). In 22 drives the camera had been recording for more than a minute (up to hours) before the table zero; that footage is removed because the rule is applied to the video clock, not to the segment numbers.
  4. Each kept segment's start time was located on the device clock from the MPEG-TS presentation timestamps and checked against its nominal position ((N - k_anchor) * 60 s, within 45 s of the drive's anchor phase) and against the GPS rows labelled with the same segment. Segments that fail (misnumbered, duplicated or foreign files) are not released and appear as missing rows; 31 drives had such segments anywhere in the recording.
  5. The 33-bit MPEG-TS timestamp counter wraps after 26.5 hours of device uptime; 8 released drive(s) contain such a wrap between kept segments. The unwrapped clock was used for alignment and the per-segment mp4 timelines are unaffected.
  6. The GPS rows satisfy 0 <= time_s <= duration_video_s and equal the CANTraj rows after the shift.

Frame times inside a segment are the mp4 presentation timestamps; t_start_s + i / 20 is an approximation only, span_s and max_frame_gap_s tell how regular each segment is (the stream copy preserves the recorded timing: 12 segments contain frame drops with a gap above 0.1 s, the largest 0.80 s). Receiver latency of the GPS fixes (0.2 to 0.7 s, estimated in CANTraj's time_alignment.json) is not compensated in the files.

Layout

VideoTraj/
β”œβ”€β”€ README.md  LICENSE  DATASHEET.md  croissant.json
β”œβ”€β”€ manifests/
β”‚   β”œβ”€β”€ drives.parquet, drives.csv       one row per CANTraj main drive (3,940 rows, 3,689 released)
β”‚   β”œβ”€β”€ segments.parquet                 one row per segment index inside each released drive's kept range (ok or missing)
β”‚   β”œβ”€β”€ camera_groups.json               nominal intrinsics per camera group
β”‚   β”œβ”€β”€ region_counts.csv                drives, drivers and hours per continent / country / admin1
β”‚   β”œβ”€β”€ splits/                          split_main, split_cross_region, split_cross_metro, long_horizon, final_test (CSV)
β”‚   β”œβ”€β”€ splits.json                      split definitions, counts and invariants
β”‚   └── CHECKSUMS.sha256                 sha256 of every file
β”œβ”€β”€ videos/driver_XXX/drive_YYY/segN.mp4  N = original segment number (seg_first .. seg_last, holes absent)
└── gps/driver_XXX/drive_YYY.csv

Column reference

manifests/drives.parquet (also drives.csv)

column meaning
driver, drive, set pseudonyms; set is always main
status ok, ok_holes (missing segments inside the kept range) or no_video (CANTraj drive without a released video; all video columns empty)
seg_first, seg_last, n_segments, n_missing, missing_segments kept range of original segment numbers, number of published segments, holes inside the range and their numbers
n_inconsistent segments of the whole recording that failed the timing check (never published)
duration_video_s, n_frames_total, bytes_total released span, frames and bytes
t_offset_cantraj_s, cantraj_duration_s, trim_head_s, trim_tail_s release zero on the CANTraj axis, length of the CANTraj table, seconds of video removed at the head and the tail on the video clock
gps_source, gps_rate_hz receiver (qcomdiag = Qualcomm modem, ublox) and median rate from CANTraj
gps_rows, gps_rows_ok, first_fix_s, last_fix_s, max_gap_s, n_gaps_gt30, fix_frac_1hz, hole_frac, path_len_km, n_pos_jumps GPS summary computed on the released window over rows with fix_ok = 1: counts, first and last fix, largest gap, gaps above 30 s, share of whole seconds with a fix, 1 - fix_frac_1hz, path length (steps implying more than 100 m/s skipped), implausible jumps between consecutive fixes of one receiver
continent, country, admin1 coarse region of the drive's median position: continent, ISO 3166-1 alpha-2 country, ISO 3166-2 first-level subdivision for US, CA and DE only (empty elsewhere)
metro_cluster anonymous integer id of a connected cluster of 0.5-degree tiles with at least 20 released drives; -1 when none
camera_group cam_a, cam_b or cam_c (see Camera groups)
op_major openpilot major.minor version of the logs (e.g. 0.9, 2026.001)
year_month month of recording (YYYY-MM); empty for 0 drives whose device clock was unusable
model_canon vehicle platform as in the DriveDNA index
anchor_verified the CANTraj time anchor was verified against the released speed (false for the unverifiable stationary drives)
flag_pts_wrap, flag_long_prestart PTS counter wrap inside the kept range; camera recording began more than 60 s before the table zero
split_main, split_cross_region, split_cross_metro, long_1h, long_2h, final_test, final_test_unseen_roads, final_test_long, calibration_region splits and cohorts (see Splits)
route_family, same_driver_max_containment, other_driver_max_containment, train_overlap CANTraj route-overlap statistics

manifests/segments.parquet

column meaning
driver, drive, seg pseudonyms and original segment number
status ok (file published) or missing (inside the kept range but absent from the recording or withheld by the timing check)
path videos/driver_XXX/drive_YYY/segN.mp4
t_start_s, t_end_s first and last frame time on the release axis (t_end_s includes the last frame's 50 ms)
n_frames, span_s, max_frame_gap_s, fps_nominal frames in the mp4, last minus first presentation time, largest step between frames, 20
width, height, codec, bytes, sha256 as staged
src_had_audio the source segment carried an audio track (removed)

gps/driver_XXX/drive_YYY.csv

The 17 CANTraj columns, unchanged except time_s (release axis, 3 decimals): time_s, msg, source, log, seg, the WGS-84 position (latitude and longitude in degrees) with altitude_m (ellipsoid height), speed_mps, bearing_deg, h_acc_m, v_acc_m, speed_acc_mps, bearing_acc_deg, flags, has_fix, fix_ok. Filter on fix_ok == 1; see the CANTraj card for the per-receiver validity rule and the meaning of each field. seg is the log segment index, which matches the mp4 segment numbers of this release.

Camera groups (nominal intrinsics)

camera_groups.json gives, per group, the declared road-camera model of the device family in its native frame and the pinhole matrix scaled to the released frame size (f_q = f_native * width_q / width_native, principal point at the frame centre). These are camera model / intrinsics groups, nominal, not per-unit calibrated: no per-vehicle calibration, mounting angle or lens distortion is released.

group device family native frame released frame f_q (px) hfov (deg) mode drives segments
cam_a 1928x1208 road-camera family (ar0231 and ox03c10 sensors on tici, tizi and mici devices) 1928x1208, f 2648 px 526x330 722.4 40.0 declared 3,181 80,076
cam_b os04c10 road camera (1344x760 binned frame) on tizi and mici devices 1344x760, f 1142 px 526x330 446.8 61.0 verify 501 13,661
cam_c comma two (neo) road camera 1164x874 1164x874, f 910 px 480x360 375.3 65.2 declared 7 205

cam_b is marked verify: the native aspect ratio does not match 526x330, so the released frames are an anisotropic scale or a crop of the sensor frame that has not been confirmed. The focal length of cam_c (comma two) is width-scaled from the declared native value; the device's own calibration is not available.

Splits

All splits are driver-disjoint; splits.json carries counts and the invariants checked at release time.

split_main role drives drivers hours >= 1 h >= 2 h
train dev 2,060 212 852.1 179 41
val dev 728 45 301.8 63 24
test final_test 309 45 196.1 38 10
few_shot_heldout few_shot_heldout 475 53 188.1 46 21
unassigned train_only 117 85 28.2 5 1
  • split_main is CANTraj's: train and val are the development folds (dev), unassigned drives may be used for training only, test is the final test (run once), and few_shot_heldout is a second held-out set, reported separately.
  • split_cross_region: drivers are assigned whole by the majority continent of their drives. train = North American drivers' train/val/unassigned drives (2,652 drives); test = drivers from other continents (312 drives in 36 countries); calibration_region = about 15 % of the drivers of each of those countries, at least one per country with three or more drivers, held back whole for per-region calibration (12 drivers, 52 drives); excluded = North American held-out drives and drives without a geocode (673).
  • split_cross_metro is copied from CANTraj and defined by drive location, not by driver: test = drives inside one held-out metropolitan area (103), excluded = drives inside a second metropolitan area that CANTraj keeps out of this split (576), train = everything else (3,010). A driver can have drives on both sides; the areas are not named.
  • long_1h / long_2h flag drives with at least 1 h / 2 h of released video (331 / 97). final_test_long = long_1h drives in test or few_shot_heldout (84), to be reported by duration bin; final_test_unseen_roads = test drives whose roads are at most 20 % contained in any other driver's roads (217).

Coverage

Drives per country (median position of the drive; admin1 adds the US state, Canadian province or German state):

continent country drives drivers hours
North America US 2,942 327 1268.6
North America CA 383 24 147.7
Asia TW 48 8 18.6
South America BR 32 5 10.9
Oceania AU 31 7 19.1
Europe DE 31 7 17.3
Asia KR 29 10 15.4
Asia AE 17 1 4.1
Europe SE 14 3 7.5
Asia TH 14 4 3.1
Europe NL 13 5 6.5
Europe SK 13 2 14.3
Europe BE 11 3 6.6
Asia MY 11 3 5.1
Europe FR 8 4 2.8
Europe IT 8 3 1.5
Europe NO 8 1 0.8
Europe RU 8 1 2.9
Asia VN 7 1 1.2
Europe CH 6 1 1.4
Asia CN 6 1 1.5
Europe GB 6 2 1.8
Europe AT 5 1 1.3
Europe CZ 5 3 0.6
Europe EE 5 1 1.0
Europe PL 4 2 0.6
Asia SA 4 2 0.2
Europe DK 3 2 0.8
Europe ES 3 1 0.6
Oceania NZ 3 1 0.6
South America AR 2 1 0.5
Asia IN 2 1 0.6
Asia KZ 2 1 0.3
Europe BG 1 1 0.2
Asia IL 1 1 0.1
Asia JO 1 1 0.1
Asia KG 1 1 0.4
Europe LT 1 1 0.1

GPS reference by receiver and rate:

receiver drives drivers hours
Qualcomm modem, about 1 Hz 2,776 250 1172.6
u-blox, 10 Hz 489 147 200.0
u-blox, about 1 Hz 424 106 193.8

Released video length per drive:

duration drives hours in test in few_shot_heldout
under 15 min 1,871 215.9 112 281
15 to 30 min 899 320.6 104 96
30 to 60 min 588 405.1 55 52
1 to 2 h 234 314.8 28 25
2 to 4 h 83 206.0 5 21
4 h and more 14 104.0 5 0

Known limitations

  • Resolution and bit rate: only the 526x330 (comma two: 480x360) qcamera stream at about 256 kbps is released. Small text, distant signs and night scenes are often unreadable at this size.
  • GPS reference: 87 % of the drives have about 1 Hz fixes, 2,776 of them from Qualcomm modem receivers without a horizontal-accuracy field; 913 drives have u-blox receivers (10 Hz where the full log existed, about 1 Hz otherwise). Receiver latency (0.2 to 0.7 s) is not compensated, and multipath in dense urban areas is present as logged. Treat the reference as accurate to a few metres, not centimetres.
  • Holes and withheld segments: 2 drives have 41 missing segments inside their kept range; segments that failed the timing check are withheld. Both appear as missing rows.
  • Frame drops: the stream copy preserves the recorded frame timing, so device-side frame drops remain visible: 12 segments have a frame gap above 0.1 s (largest 0.80 s) and 4 segments have fewer than 1,150 frames. Segments with a gap above 0.5 s: driver_079/drive_030 seg 38, driver_079/drive_030 seg 39. Segments with fewer than 1,150 frames: driver_078/drive_006 seg 40, driver_079/drive_030 seg 38, driver_079/drive_030 seg 39, driver_319/drive_011 seg 12. max_frame_gap_s and n_frames in segments.parquet locate them; no frame was dropped or added by the release process.
  • Geographic skew: 80 % of the released drives are in the United States and 18 % in Florida; the other continents together hold the cross-region test set and are small per country.
  • Camera metadata is nominal (see above), and the 7 comma two drives use a different sensor and frame size.
  • Device heterogeneity: openpilot versions from 0.7 to 2026 releases, several device generations and 208 vehicle platforms.
  • Alignment residue: segment starts are exact on the device clock, but the CANTraj anchor itself carries the uncertainty documented in the CANTraj card (anchor_verified), and 56 released drives could not be verified against the vehicle speed.
  • No CAN in this repository: speed, steering and IMU are in CANTraj under the same pseudonyms and axis (time_s + t_offset_cantraj_s).

Privacy

  • The drives were collected from community drivers with informed consent and compensation, following the source platform's terms (as stated for DriveDNA). Drivers and drives appear only as the sequential pseudonyms of DriveDNA; the private mapping is held by the authors.
  • Trimming: the first and last two minutes of every drive are removed on both clocks, so the origin and the destination of a trip are not in the video or in the GPS file. Two minutes of driving are roughly one kilometre, so the released span can still start near an origin; this is the same residual risk as the untrimmed first segments that DriveDNA already publishes under the same gate, and this release adds alignment and length rather than new exposure. GPS coordinates inside the kept window are exact, which trajectory evaluation and map matching need; for that reason the repository is gated and research use only.
  • No absolute time (only the month), no audio, no cabin camera, no vehicle or device identifier. Region labels are coarse (country, first-level subdivision for three countries, anonymous metro cluster ids).
  • Road video shows public roads and may contain other vehicles' licence plates and pedestrians at low resolution; do not attempt to identify people or vehicles.
  • Prohibited: re-identification of drivers; inference of home, work or other personal locations; insurance, employment, credit or law-enforcement scoring of individuals; redistribution outside the licence.
  • Takedown: a takedown contact (haozhou1@usf.edu, as on the DriveDNA and CANTraj cards) allows any driver to request removal from future versions; users agree to delete removed drives from local copies on notice (licence clause 5).

Quick start

pip install -U huggingface_hub pandas pyarrow av
hf download HenryYHW/VideoTraj --repo-type dataset --include "manifests/*" --local-dir VideoTraj
hf download HenryYHW/VideoTraj --repo-type dataset --include "videos/driver_027/drive_002/*" "gps/driver_027/drive_002.csv" --local-dir VideoTraj
import av, pandas as pd

drives = pd.read_parquet("VideoTraj/manifests/drives.parquet")
segs = pd.read_parquet("VideoTraj/manifests/segments.parquet")
row = drives[(drives.driver == "driver_027") & (drives.drive == "drive_002")].iloc[0]
gps = pd.read_csv("VideoTraj/gps/driver_027/drive_002.csv")
gps = gps[gps.fix_ok == 1]

# frames of one segment with their time on the release axis
s = segs[(segs.driver == row.driver) & (segs.drive == row.drive) & (segs.status == "ok")].iloc[0]
with av.open("VideoTraj/" + s.path) as c:
    st = c.streams.video[0]
    for frame in c.decode(st):
        t = s.t_start_s + float(frame.pts * st.time_base)      # seconds since the first released frame
        img = frame.to_ndarray(format="rgb24")                  # 330 x 526 x 3
        break

# the GPS fix nearest to that frame, and the same instant on the CANTraj / DriveDNA axis
fix = gps.iloc[(gps.time_s - t).abs().argmin()]
t_cantraj = t + row.t_offset_cantraj_s

With the datasets library the CSV manifests and the GPS files load as configs (drives, split_main, ..., gps); the videos are plain files to download.

License

Released under the DriveDNA research license (data): research use only, no re-identification, no redistribution, takedown on request, attribution. The full text in LICENSE is the data licence published with the DriveDNA Hub card. Access is gated and approved manually.

Citation

VideoTraj paper: citation to follow (this card will be updated when the paper is available). Until then, refer to this release as VideoTraj-Long v0.1 (Yuhang Wang, Lingyao Li and Hao Zhou, University of South Florida and University of Arizona, 2026; the authors of DriveDNA and CANTraj) and cite the DriveDNA paper for the underlying corpus:

@article{drivedna2026,
  title   = {DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and
             Benchmark for Driving Style Identification},
  author  = {Wang, Yuhang and Li, Lingyao and Zhou, Hao},
  journal = {arXiv preprint arXiv:2607.23822},
  year    = {2026}
}

The CANTraj release provides the GPS reference and the splits; its paper citation is also to follow (see the CANTraj card).

Links

VideoTraj-Long v0.1, University of South Florida and University of Arizona, 2026, card generated 2026-11-09
Downloads last month
2,474

Paper for HenryYHW/VideoTraj