README.md

metadata

license: cc-by-4.0
task_categories:
  - video-classification
  - object-detection
tags:
  - surveillance
  - fall-detection
  - sentinel
  - video-intelligence

Le2i Sentinel Test Frames

Extracted frames from the Le2i Fall Detection Dataset for benchmarking the Sentinel video intelligence pipeline.

Contents

• 130 annotated videos (99 falls, 31 normals)
• 3 frames per video (390 total)
• 4 environments: Coffee_room_01, Coffee_room_02, Home_01, Home_02
• Ground truth in metadata/ground_truth.json

Source

Le2i Fall Detection Dataset (University of Burgundy)

• Resolution: 320x240 @ 25fps
• Citation: Charfi et al., "Optimised spatio-temporal descriptors for real-time fall detection", JEI 2013

Usage

from huggingface_hub import snapshot_download
snapshot_download("PixelML/le2i-sentinel-frames", local_dir="le2i_frames")

Benchmark Results (Sentinel V03 on Mistral Small 3.2)

• Fall detection: F1=0.889, P=0.988, FPR=0.032
• Cosmos-Embed1 alert gate: 9/10 correct