mono-aeb

A detector-agnostic monocular AEB (Autonomous Emergency Braking) / FCW (Forward Collision Warning) decision chain: detection → tracking → in-path filtering → ranging → TTC → risk decision.

The two core strengths are calibration-free scale TTC — which relies only on the rate of change of a target's image size and does not depend on camera intrinsics or distortion calibration — and speed-adaptive risk thresholds, where the AEB trigger point shifts with ego speed. The detector is pluggable: the YOLOv8 backend (yolo extra) and the high-accuracy D-FINE backend (dfine extra) both implement the same BaseDetector interface, and the decision chain consumes only [x1, y1, x2, y2, score, class].

中文说明见 README.zh-CN.md。

Features

• Detector-agnostic contract — the decision chain is decoupled from any specific detector and works with the 7 AEB classes (person / rider / car / truck / bus / bike / motor).
• Calibration-free scale TTC — τ = s / ṡ, estimated via a Theil–Sen robust fit of (t, 1/s); no camera calibration required.
• Distance TTC + fusion — a ground-plane ranging model recovers closing speed, fused with scale TTC by taking the minimum, with a persist_k frame hold to suppress jitter.
• Speed-adaptive thresholds — τ_AEB(v) = t_react + v / (2·a_max), with fixed lead margins for FCW / ATTENTION on top; the whole decision gradient shifts with ego speed.
• Friction-aware thresholds (opt-in) — set RiskThresholds.friction_mu and the same formula uses a measured a_eff(μ) table (aeb/friction.py) instead of the dry-road a_max, so low-friction surfaces trigger earlier; the default (None) keeps the fixed-a_max behaviour bit-for-bit.
• Per-video self-calibration — jointly estimates the ranging scale f·H and the horizon horizon_y from GPS ego speed + calibration-free scale TTC, removing the per-video camera-parameter inconsistency in BDD100K.
• Bundled ByteTrack — the tracker is vendored into the package (aeb/tracker/vendor/bytetrack/) and distributed with the wheel; no separate installation needed.

Pipeline

video frame → detection (YOLOv8 / D-FINE)
            → tracking (ByteTrack)
            → in-path filtering (fixed trapezoid ROI, ego-path corridor projection)
            → ranging (ground-plane model, D = f·H / (y_bottom − horizon))
            → TTC (scale τ and distance τ fused via min)
            → decision (state machine NORMAL → ATTENTION → FCW → AEB_WARNING)

The risk state machine only escalates while a target is inside the ego-path corridor (in_path = True); targets outside the path stay at NORMAL, avoiding false triggers from adjacent-lane vehicles.

Installation

Requires Python ≥ 3.10.

From PyPI:

pip install mono-aeb

From source (for development):

pip install -e .

Runtime dependencies:

About lap: lap is the linear-assignment solver used by ByteTrack matching. If no lap wheel exists for your platform/Python, install lapx instead — it is an alternative distribution of lap that ships a module named lap:

pip install lapx

Note: this project does not fall back in code. When lap is missing, installing lapx as above is enough (it provides the module name lap, so no code changes are needed).

ByteTrack itself is vendored as a minimal subset (aeb/tracker/vendor/bytetrack/, MIT License) and installed with the wheel, so both pip install -e . and a regular pip install work without fetching the upstream repository. To use an external/full ByteTrack instead:

If pip install -e . aborts because no lap wheel is available for your platform, install lapx first, then install this package while skipping dependency resolution:

pip install lapx
pip install -e . --no-deps

Optional high-accuracy backend D-FINE additionally requires the PyTorch ecosystem:

pip install -e ".[dfine]"   # torch / torchvision / pillow

About Windows: torch requires the Microsoft Visual C++ runtime; on some machines (especially lab/clean-room environments) its absence shows up as OSError: [WinError 126] ... c10.dll failing to load. Installing the VC++ Redistributable (x64) once resolves it. (The --demo zero-data demo does not import torch, so it does not need this.)

On Windows consoles whose code page is not UTF-8, non-ASCII text (for example in your own file paths, or in messages coming from third-party libraries) can still raise UnicodeEncodeError; run set PYTHONUTF8=1 first. Linux/macOS and UTF-8 consoles are unaffected.

Quick start

Zero-data demo (no video / no weights)

After pip install mono-aeb (no video, no weights, no network):

mono-aeb demo

The same command without the console script is python -m aeb demo.

This command drives a synthetic detector that generates a target closing at constant speed, and runs the full detection → tracking → ranging → TTC → decision chain, printing state changes frame by frame:

[quickstart] ego speed = 12 m/s, AEB threshold = 1.37s
[quickstart] target closing at a constant 10 m/s; printing the risk state per frame:
  frame   0: NORMAL     d= 45.0m  ttc=  infs
  frame  39: ATTENTION  d= 32.0m  ttc= 3.36s
  frame  69: FCW        d= 22.0m  ttc= 2.36s
  frame  99: AEB        d= 12.1m  ttc= 1.36s
[quickstart] done, 120 frames

Available options: --out video.mp4 (write a visualization video), --show (live window), --frames / --fps / --v-close. The demo mode does not require ultralytics / torch; it depends only on numpy / scipy / opencv-python / lap.

Dependency self-check:

mono-aeb check

(From a source checkout without installing, the equivalent commands are python examples/quickstart.py --demo and python examples/quickstart.py --check.)

Real video

mono-aeb run --source video.mp4 --coco --out out.mp4

(From a source checkout: python examples/demo_aeb.py --source video.mp4 --coco --out out.mp4.)

• --coco: maps COCO classes to the 7 AEB classes when using COCO-pretrained weights (e.g. yolov8n.pt); omit it when using AEB-finetuned weights.
• --ego-speed: ego speed (m/s); when omitted, GPS speed is read from BDD100K samples-1k/info/*.json automatically, falling back to 12 m/s when unavailable.
• --detector dfine: switch to the D-FINE backend (see below).

You may also pass an image directory via --source ./frames_dir.

Detector backends

YOLOv8 (optional backend, yolo extra)

pip install -e ".[yolo]"    # ultralytics (AGPL-3.0)

from aeb.detectors import YOLOv8Detector, COCO_TO_AEB

det = YOLOv8Detector(
    "yolov8n.pt", conf=0.25, device="0", class_map=COCO_TO_AEB
)  # class_map=None for finetuned weights

Optional high-accuracy backend: D-FINE

D-FINE depends on its upstream repository, located via environment variables to avoid hardcoding local paths:

export DFINE_ROOT=/path/to/D-FINE
export MONO_AEB_DFINE_WEIGHTS=/path/to/best_stg1.pth
mono-aeb run --source video.mp4 --detector dfine

(From a source checkout the same flags work through the shim: python examples/demo_aeb.py --source video.mp4 --detector dfine.)

Or pass the arguments directly in code:

```python
from aeb.detectors import DFineDetector

det = DFineDetector(weights="/path/to/best_stg1.pth")

The config argument defaults to <DFINE_ROOT>/configs/dfine/dfine_hgnetv2_m_aeb.yml; pass config=... to use another file.

Note: D-FINE's finetuned weights are trained on BDD100K and are subject to that dataset's license (see below); they are not distributed with this repository.

Weights & data

Compliance note: this repository does not contain any model weights or BDD100K videos/annotations.

• COCO-pretrained YOLOv8 weights (e.g. yolov8n.pt) are provided by ultralytics and obtained under ultralytics' license.
• AEB-finetuned weights (YOLOv8 or D-FINE) are finetuned on BDD100K, whose license is "academic/non-commercial use only, no redistribution of raw data". These weights are therefore not published here; they are distributed on a restricted basis (on request). With legitimate data and license access they can be reproduced by training the upstream detector on a 7-class AEB dataset, then pointing the backend at the resulting config and weights (DFINE_ROOT / MONO_AEB_DFINE_WEIGHTS for D-FINE). This repository does not ship a training configuration.

Project structure

mono-aeb/
├── aeb/
│   ├── cli.py              # console entry point: mono-aeb demo / check / run / calib
│   ├── __main__.py         # python -m aeb
│   ├── config.py           # camera / ego-path / risk thresholds / detection & tracking params
│   ├── friction.py         # measured μ → a_eff table for friction-aware thresholds
│   ├── types.py            # Detection / Track / RiskFeature / RiskLevel contracts
│   ├── pipeline.py         # AEBPipeline main chain
│   ├── calibration.py      # per-video self-calibration (joint f·H & horizon estimation)
│   ├── ego_speed.py        # GPS speed from BDD100K info/*.json
│   ├── imageio.py          # unicode-safe image read/write (cv2 fails silently on non-ASCII paths)
│   ├── detectors/          # BaseDetector + YOLOv8 + D-FINE
│   ├── tracker/            # ByteTrack adapter + cython_bbox numpy shim
│   │   └── vendor/bytetrack/   # vendored ByteTrack (MIT, minimal subset, shipped with wheel)
│   ├── distance/           # ground-plane ranging
│   ├── ttc/                # scale TTC / distance TTC / fusion / track history
│   ├── in_path/            # fixed trapezoid ROI (ego-path corridor)
│   └── decision/           # risk state machine
├── examples/               # thin shims for checkout use; the logic lives in aeb/cli.py
│   ├── quickstart.py       # → mono-aeb demo / check
│   ├── demo_aeb.py         # → mono-aeb run
│   └── calibrate_video.py  # → mono-aeb calib
├── tests/                  # zero-data test suite (no video / weights / network)
├── benchmarks/             # synthetic tracking benchmark, bbox_overlaps shim cost
├── .github/workflows/      # CI (tests / lint) and Release (tag → GitHub Release, PyPI)
├── Dockerfile              # python:3.12-slim + dev extra, runs the zero-data suite
├── CONTRIBUTING.md
├── ROADMAP.md
├── CHANGELOG.md
├── pyproject.toml
├── LICENSE
├── README.md               # this file (English, primary)
└── README.zh-CN.md         # Chinese counterpart

Core principles

Ranging (ground-plane model)

Assuming a flat ground plane with camera height H and focal length f (which enter ranging only through their product f·H, a scale degeneracy):

$$ D = \frac{f \cdot H}{y_{bottom} - y_{horizon}} $$

f·H and horizon_y are calibrated per video (see below), not from global intrinsics.

Calibration-free scale TTC

For a target's image size s (box height or width), its rate of change is related to closing speed, independent of camera calibration:

$$ \tau_{scale} = \frac{s}{\dot{s}} $$

In implementation, a Theil–Sen robust linear fit is applied to (t, 1/s), yielding slope b, so τ_scale = −1 / (b · s_now). Theil–Sen takes the median of pairwise slopes and has a breakdown point of about 29% (1 − 1/√2), robust to a significant fraction of outliers — better than ordinary least squares, whose breakdown point is 0.

Distance TTC and fusion

From the ranging sequence (t, D), fit the closing speed v_close = −slope, then τ_distance = D / v_close. The two sources are fused by taking the more conservative one:

$$ \tau_{fused} = \min(\tau_{scale}, \tau_{distance}) $$

with a persist_k frame hold to suppress single-frame jitter.

Risk state machine

NORMAL(0) → ATTENTION(1) → FCW(2) → AEB_WARNING(3)

The AEB trigger threshold adapts to ego speed (t_react reaction time, a_max emergency-braking max deceleration):

$$ \tau_{AEB}(v) = t_{react} + \frac{v}{2 \cdot a_{max}} $$

FCW / ATTENTION add fixed lead margins on top (defaults +1.0s / +2.0s). Escalation is immediate; de-escalation has hysteresis to avoid oscillation around thresholds. Required deceleration (reserved, not currently wired into the state machine):

$$ a_{req} = \frac{v_{close}^2}{2 \cdot (D - d_{safe})} $$

Defaults are in aeb/config.py (t_react=0.5, a_max=6.86 m/s², d_safe=5.0 m, history_len=10, persist_k=3, fps=30).

With RiskThresholds.friction_mu set, a_max is replaced by a measured effective deceleration a_eff(μ) from aeb/friction.py — 10 measured points over μ ∈ [0.4, 3.5] (measurement protocol in that module's docstring). Values outside the range clamp to the nearest endpoint (no extrapolation) and the result is capped at a_max, so high-friction surfaces behave exactly like the fixed-a_max formula:

from aeb.config import AEBConfig

cfg = AEBConfig()  # fixed dry-road a_max
cfg.risk.friction_mu = 0.6  # low-friction surface: a_eff(0.6) = 3.566 m/s², triggers earlier

Per-video calibration

BDD100K is crowdsourced, so cameras differ per video and there are no unified intrinsics. This repository provides two calibrations:

1. Ego speed + scale TTC (primary entry point): for approximately stationary targets D = v_ego · τ_scale, then a Theil–Sen line fit over (1/D, y_bottom) takes its intercept as horizon_y; f·H is not the fitted slope but a per-track median of D·(y_bottom − horizon) (the Theil–Sen slope is more sensitive to tracking-smoothing lag) (aeb/calibration.py: estimate_ground_plane).
2. Known target width (reserved): CameraConfig.calibrate_from_known_width() (currently no call site, kept as a reserved interface) recovers camera height from how pixel width changes with (y_bottom − horizon).

CLI:

mono-aeb calib --stem 0571873b-faf718b2
# or
mono-aeb calib --video /path/to/samples-1k/videos/xxxx.mov --max-frames 200

(From a source checkout: python examples/calibrate_video.py takes the identical flags.) --stem mode locates the video directory via the BDD100K_VIDEO_DIR environment variable (defaults to the current directory).

Development & testing

Zero-data test suite: no video, detection weights, or network needed — only the core dependencies (numpy / scipy / opencv-python / lap).

pip install -e ".[dev]"     # or: pip install pytest pytest-cov (--cov needs pytest-cov)
pytest -q --cov=aeb --cov-report=term-missing

Coverage: the zero-data suite covers 69% of aeb/ overall (re-measured 2026-10-16, incl. the decision gate; 1564 statements) — 72% over the decision chain alone, i.e. excluding the two CLI entry-point files (aeb/cli.py 56%: demo / check / parser / --version are driven in-process by tests/test_cli_module.py, while its run / calib paths need the optional backends; aeb/__main__.py runs only inside the subprocess smoke tests, so it reports 0% in-process). The decision chain itself is well covered — pipeline.py 100%, in_path/fixed_roi.py 96%, tracker/bytetrack.py 96%, ttc/* 84–92%, calibration.py 85%, friction.py 100% — while the detector implementation modules (detectors/dfine_detector.py, detectors/yolov8_detector.py) and ego_speed.py report 0% because they need the optional detector backends and real GPS data. CI prints the per-module table on every run (fully covered files are skipped).

Lint / formatting: ruff, configured in pyproject.toml (the vendored ByteTrack subset is excluded; CJK punctuation in comments is allowed), plus a pre-commit hook set in .pre-commit-config.yaml. Run pip install pre-commit && pre-commit install to enable it locally; a dedicated CI job runs ruff check and ruff format --check.

CI (.github/workflows/ci.yml) runs the above tests and the zero-data demo on Ubuntu / Windows × Python 3.10 / 3.12, plus a dedicated lint job; a further job runs pip install -e ".[yolo,dev]", validating full dependency resolution including the optional detector backend (yolo extra).

Container

A Dockerfile builds the same environment in one step — python:3.12-slim plus the dev extra, and no compiler, because every core dependency has a manylinux wheel:

docker build -t mono-aeb .
docker run --rm mono-aeb                                          # pytest
docker run --rm mono-aeb python examples/quickstart.py --check    # zero-data demo

The image is built from this directory and is not published to a registry. The default command is the zero-data test suite, and the container needs no network access at test time.

License

This project's code is released under Apache-2.0; see LICENSE.

Third-party components & acknowledgments

License note: ultralytics is AGPL-3.0, which imposes copyleft obligations on derivative works and network-service scenarios. If your distribution needs to avoid those obligations, use the Apache-2.0 D-FINE detector instead (--detector dfine). This repository's own code is Apache-2.0.