400 hours of in-cabin driver motion · 360 drivers · 9,010 sequences · 680,082 forecasting windows 133 whole-body keypoints @ 10 Hz · synchronized CAN & exterior context · 5 camera-view types

🎬 What does it look like?

Every sequence in DriveMotion is released as a privacy-reduced skeleton motion video plus a standardized keypoint tensor — the driver's behavior is preserved, appearance identity is not.

Fleet (BATON)	In-the-wild web	AIDE (semantic labels)

continuous routes, CAN, head pose	varied viewpoints & visibility	behavior / emotion annotations

🧭 Why DriveMotion?

Driver-monitoring datasets are built for recognizing actions from short clips. Human-motion forecasting benchmarks live in labs or on sidewalks. Neither covers the question an in-cabin system actually faces: what will the driver's body do in the next few seconds?

DriveMotion standardizes three heterogeneous sources into one motion representation and one forecasting protocol:

Source	Role	Sequences	Hours	What it brings
BATON fleet	temporal backbone	1,347 routes	320 h	continuous minutes-to-hours routes, time-aligned CAN, device-grounded head pose, road-camera context
Web corpus	observation diversity	4,765 spans	78 h	front / side / back viewpoints, five visibility levels, creator-diverse cabins
AIDE (re-extracted)	semantics	2,898 clips	2.4 h	behavior & emotion labels in the same sequence format

All sources pass through one extraction trunk (RTMW whole-body pose, resampling to 10 Hz, normalization, validity estimation, quality scoring) and differ only in how the driver is found:

📦 Unified representation

One sequence = one .npz + one .meta.json, on a fixed 10 Hz grid:

Field	Shape	Meaning
`t` / `t_actual`	`[N]`	grid time / actual source-video time (s)
`kpts`	`[N, 133, 3]`	COCO-WholeBody layout `(x, y, score)`, normalized image coords; 127 slots active in-cabin
`mask`	`[N, 133]`	per-joint validity — occlusion is explicit, never imputed
`head`	`[N, 3]`	yaw / pitch / roll (deg); device-derived on BATON, vision-based elsewhere
`can`	`[N, 4]`	speed, steering, throttle, brake (BATON)
`part_valid`	`[N, 6]`	per-part coverage flags
`quality`	`[N]`	per-frame extraction quality score

meta.json records source, view type, visibility level, driver id, native fps/resolution, and absolute source timestamps — the full provenance of every frame.

import numpy as np, json

d = np.load("motion/web/<seq>.npz")
meta = json.load(open("motion/web/<seq>.meta.json"))
kpts, mask = d["kpts"], d["mask"]            # [N,133,3], [N,133]
print(meta["view"], meta["driver_visibility"], kpts.shape)

🎯 The forecasting benchmark

Task: observe 8 s of driver motion → predict the next 4 s of keypoint trajectories and head pose (10 Hz), in a canonical torso frame. Exterior scene features (2 Hz frozen ResNet-50 embeddings of the road view) are an optional input. CAN is never an input — it is used only offline to build evaluation windows.

Naturalistic driving is dominated by stillness, so uniform evaluation mostly scores "nothing happens". DriveMotion therefore anchors its primary protocol on vehicle-dynamics transitions (76,026 maneuver initiations mined from CAN):

Protocol A — dynamics-anchored forecasting: pre-maneuver / post-maneuver / stable-control strata (6,138 / 6,238 / 4,262 test windows). Arm motion in pre-maneuver windows is 3.4× that of matched stable driving.
Protocol B — robustness under observation shift: train on the fleet, evaluate on held-out web creators across viewpoints and visibility levels.
Metrics: MPJPE@4s on 23 cross-view-stable points, plus Part-State F1@2s — does the head / torso / each arm stay still, move a little, or move a lot?
Identity-disjoint splits, fixed hashed evaluation subsets, frozen state thresholds — all released with the toolkit.

Reference results (dynamics-anchored test set, 16,638 windows)

Model	MPJPE@4s ↓	Part-State F1@2s ↑
Zero-motion (persistence)	7.75	0.215
GRU seq2seq	6.83	0.275
siMLPe	6.85	0.227
Transformer ED	6.75	0.282
Transformer ED (+ctx, enriched)	6.95	0.309
Transformer-L	6.63	0.287
Transformer-XL	6.62	0.284
CVAE	6.84	0.257
DDPM (+ctx, enriched)	8.86	0.299
AR motion-token LM (enriched)	8.16	0.458
LLM backbone (Llama-3B)	8.24	0.457

Two findings the benchmark is designed to expose: forecasting skill concentrates in motion-active, maneuver-related intervals, and geometric accuracy and behavioral-state anticipation favor different model families — coordinates alone don't tell you whether a hand is about to move.

🤖 Can models actually predict driver motion?

Watch a forecaster work on anchored pre-maneuver windows — the model sees 8 s of motion history (plus road context) and rolls out 4 s into the future:

_{Left: ground truth (blue trails = nose/wrists). Right: Transformer (+ctx, enriched)
prediction (orange trails) over the faint ground-truth ghost.}

Stochastic models produce diverse futures on the same observation:

🗂️ Repository layout

DriveMotion/
├── assets/                # card figures & GIFs
├── motion/                # ⬆ uploading — npz + meta.json per sequence
│   ├── baton/  ├── web/  └── aide/
├── render/                # ⬆ uploading — skeleton motion videos (*.motion.mp4)
├── events/                # CAN maneuver event bank (per-route parquet)
├── benchmark/             # manifest, splits, anchored windows, thresholds,
│                          # fixed subsets, exterior-context features, caches
├── code/                  # loader + full benchmark reference implementation
└── demo/                  # runnable demo script + sample sequences

🚧 Upload in progress — the full data payload (~330 GB) is being pushed in stages. The card, demos, and benchmark definitions land first; motion tensors and skeleton videos follow.

🔒 Privacy & license

The released visual modality is skeleton motion video — behavior is preserved while appearance-based identity is substantially reduced. No identity-preserving crops of drivers are distributed for the web corpus.
Every web curation decision is logged: the funnel ledger (per-stage rejection causes and gate values) ships with the dataset, and takedown requests are honored.
Skeleton-derived artifacts, annotations, metadata, and code produced by DriveMotion: CC BY 4.0. Source-derived artifacts retain the licensing terms of their respective sources; aligned context-view clips are released separately under a research-only license.
Splits are grouped by recording device / creator channel so no driver crosses a split boundary.

📖 Citation

A technical report describing DriveMotion is in preparation — citation information will be added here.

✉️ Contact

Open a discussion on this repository for questions, issues, or takedown requests.

Downloads last month: 134

Total file size:

261 GB