Agri-Human Dataset Benchmark

This repository is a benchmark workspace for agricultural human detection and tracking, built to support safer human-robot interaction in field robotics.

The core idea is simple: if agricultural robots are expected to work around people, they need perception systems that can detect, track, and eventually fuse observations of humans reliably across difficult real-world conditions such as occlusion, clutter, motion, wide-angle optics, and changing light. This repository exists to make those comparisons repeatable.

Today, the implemented code focuses on 2D detection and 2D tracking. The broader benchmark direction is multimodal: RGB, fisheye, LiDAR point clouds, rosbag-derived streams, and future 3D or fusion pipelines can be added under the same benchmark philosophy.

Why This Repo Exists

Agricultural robotics operates in environments where standard benchmark assumptions often break down:

• people move unpredictably around machines
• sensing conditions vary across rows, canopies, footpaths, and work zones
• camera geometry may be conventional RGB, stereo, or fisheye
• safety depends on robust perception, not just high scores on generic urban datasets

This repository is intended to help answer questions such as:

• How well do person detectors trained on one agricultural domain transfer to another?
• How much performance changes between RGB and fisheye imagery?
• Which trackers are most stable when detections are noisy or sparse?
• How should future 3D and multimodal benchmarks be organised so they remain comparable?

What Is Implemented Now

2d-detection

Training, evaluation, transfer evaluation, and result summarisation for 2D person detection benchmarks.

• benchmarks/ultralytics: YOLO-based benchmark runner and detection export pipeline
• benchmarks/mmdetection: MMDetection-based benchmark runner for Faster R-CNN style experiments
• reports/benchmarks: committed benchmark summaries and selected run artifacts
• notebooks: analysis notebooks

2d-tracking

Tracker benchmarking and MOT-style evaluation for detections exported from the 2D detection workspace.

• benchmarks/boxmot: BoxMOT wrappers for ByteTrack, BoTSORT, StrongSORT, OCSORT, DeepOCSORT, and BoostTrack
• benchmarks/deepsort: local DeepSORT implementation with frozen-graph ReID support
• benchmarks/norfair: Norfair-based tracking benchmark
• common/mot: shared ground-truth conversion and MOT evaluation utilities
• reports: per-run tracking outputs and summary metrics

Benchmark Scope

The repository is designed around benchmark comparison, not around a single model implementation.

Current benchmark patterns in the codebase include:

• in-domain train and evaluate
• zero-shot or pretrained evaluation without fine-tuning
• cross-domain transfer evaluation
• export of detector outputs into a common JSON format for downstream tracking
• MOT-format evaluation of tracker outputs

The current configs already reflect several image domains, including zedrgb, fisheye, fieldsafepedestrian, kitti_filtered, and coco2017_filtered.

High-Level Workflow

1. Train or evaluate 2D detectors in 2d-detection.
2. Export frame-wise detections to the repository JSON format.
3. Run tracker benchmarks in 2d-tracking.
4. Convert project annotations to MOT ground truth and evaluate tracker outputs.
5. Compare reports across datasets, models, and sensing conditions.

Repository Layout

.
|-- 2d-detection/
|   |-- benchmarks/
|   |-- notebooks/
|   `-- reports/
|-- 2d-tracking/
|   |-- benchmarks/
|   |-- common/
|   `-- reports/
`-- LICENSE

Data And Paths

The repository does not package the full datasets. Most configs point at local or cluster-mounted dataset paths such as /workspace/data/..., which you should adapt to your own environment.

Path conventions differ slightly by subsystem:

• Ultralytics benchmark report paths are resolved from 2d-detection
• MMDetection benchmark YAML paths are typically resolved from the repository root
• tracking configs resolve relative paths from 2d-tracking

Each workspace README explains its own path rules in more detail.

MMDetection Dependency

The MMDetection benchmark code in this repository is designed to use the external mmdet Python package rather than a vendored framework copy. Local benchmark logic and dataset-specific configs stay under 2d-detection/benchmarks/mmdetection/.

If framework-level MMDetection changes are ever required, they should live in an external fork of open-mmlab/mmdetection and be referenced as a dependency from environment setup documentation or requirements.txt, not copied into this repository.

Intended Direction

This repository should be read as both:

• a working benchmark implementation for current 2D perception experiments
• a foundation for publishing and extending an agricultural human-safety dataset benchmark across more sensing modalities

That future expansion can include:

• 3D detection
• 3D tracking
• multimodal fusion
• LiDAR and point-cloud benchmarks
• rosbag-to-benchmark conversion utilities
• cross-modal safety analysis for human-robot interaction

License

The repository's benchmark code and documentation are licensed under Apache License 2.0. See LICENSE.

Additional attribution information is recorded in NOTICE.

If third-party code is added to the repository in the future, its original license and notices should be preserved.