Zero-shot fall detection: YOLO-World v2 + RelateAnything, ONNX/OpenVINO for edge CPU and PyTorch for GPU
Python
0
2 commits
updated Oct 7, 2026
Zero-shot fall detection. YOLO-World v2 (YOLOv8) detects people and the floor,
and RelateAnything scores relations such as person — lying on — floor. A
per-person temporal filter then confirms the fall. See
docs/architecture.md for the full design.
bash setup.sh # conda env "Aeon_fall_detection" (Python 3.12)
conda activate Aeon_fall_detection
python export_onnx.py --openvino # once: ONNX + OpenVINO artifacts (~300 MB)
python app.py --mode onnx_cpu # Edge CPU (i7-1260P) -> http://127.0.0.1:7860
python app.py --mode torch_gpu --host 0.0.0.0 # RTX 4090 / 5090
python pipeline.py image.jpg --mode onnx_cpu --output out.jpg
python pipeline.py clip.mp4 --mode torch_gpu --output out.mp4
docker compose up -d --build
docker compose --profile export run --rm export # artifacts for edge devices
pytest -q tests
python generate_docx_report.py --out-dir docs --markdown docs/Kien_truc_He_thong_Aeon_Fall_Detection.md
objects: person, floor, bed, sofa
fall: lying on, laying on, resting on, sitting on
upright: standing on, walking on, walking past, standing beside
predicate_bank.npz. Phrases not in the
bank are reported in the alert log. To add one, re-export with
python export_onnx.py --openvino --predicates "collapsed on".bed or sofa suppresses alarms for people
resting on it.requires-python >= 3.12.Copyright (C) 2026 Anh-Nam-97
This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. It is distributed WITHOUT ANY WARRANTY; see LICENSE for the full text.
The project is AGPL-3.0 because it builds on Ultralytics (YOLO-World) and RelateAnything, both AGPL-3.0. If you run a modified version as a network service (for example the Gradio UI), section 13 requires offering its users the corresponding source code; the UI header links to this repository for that purpose.
Zero-shot fall detection: YOLO-World v2 + RelateAnything, ONNX/OpenVINO for edge CPU and PyTorch for GPU
Python
0
2 commits
updated Oct 7, 2026
Zero-shot fall detection. YOLO-World v2 (YOLOv8) detects people and the floor,
and RelateAnything scores relations such as person — lying on — floor. A
per-person temporal filter then confirms the fall. See
docs/architecture.md for the full design.
bash setup.sh # conda env "Aeon_fall_detection" (Python 3.12)
conda activate Aeon_fall_detection
python export_onnx.py --openvino # once: ONNX + OpenVINO artifacts (~300 MB)
python app.py --mode onnx_cpu # Edge CPU (i7-1260P) -> http://127.0.0.1:7860
python app.py --mode torch_gpu --host 0.0.0.0 # RTX 4090 / 5090
python pipeline.py image.jpg --mode onnx_cpu --output out.jpg
python pipeline.py clip.mp4 --mode torch_gpu --output out.mp4
docker compose up -d --build
docker compose --profile export run --rm export # artifacts for edge devices
pytest -q tests
python generate_docx_report.py --out-dir docs --markdown docs/Kien_truc_He_thong_Aeon_Fall_Detection.md
objects: person, floor, bed, sofa
fall: lying on, laying on, resting on, sitting on
upright: standing on, walking on, walking past, standing beside
predicate_bank.npz. Phrases not in the
bank are reported in the alert log. To add one, re-export with
python export_onnx.py --openvino --predicates "collapsed on".bed or sofa suppresses alarms for people
resting on it.requires-python >= 3.12.Copyright (C) 2026 Anh-Nam-97
This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. It is distributed WITHOUT ANY WARRANTY; see LICENSE for the full text.
The project is AGPL-3.0 because it builds on Ultralytics (YOLO-World) and RelateAnything, both AGPL-3.0. If you run a modified version as a network service (for example the Gradio UI), section 13 requires offering its users the corresponding source code; the UI header links to this repository for that purpose.