A full fine-tune of Qwen3.5-2B (Qwen's native image-text-to-text architecture) that detects human falls, being down, and distress from a short strip of consecutive video frames. This is a sibling model to edge-fall-vlm-2.2b (SmolVLM2-2.2B), chosen for capability over edge-device footprint — see the honest trade-off below.

Code & full write-up: https://github.com/vieenrose/edge-fall-vlm

What it does

Input: a few consecutive frames (oldest→newest). Output: JSON, e.g. {"posture":"horizontal-on-floor","status":"down","confidence":0.9,"person_down":true,"n_people":1} where status ∈ {down, distress, normal} (3-class down3 scheme).

Why this model exists (honest story)

edge-fall-vlm-2.2b (SmolVLM2-2.2B) has strong recall on our in-the-wild benchmark, but a real user-submitted clip (an elderly person falling from a bent/reaching position near a chair, in a cluttered, low-light room) was missed by every variant of it we tried — including three separate rounds of synthetic-data improvement (more fall-type coverage, procedural occlusion, and real photo-textured occlusion + indoor HDRI lighting). None of those closed the gap. Testing a different base architecture on the exact same clip did: Qwen3.5-2B, fine-tuned with the same recipe/data as the SmolVLM2 model (no extra synthetic enrichment), correctly detected the fall — the only model in the investigation to do so, verified across multiple frame windows spanning the actual fall event.

Results (OOPS in-the-wild benchmark, n=210 combined val+test — the real trade-off)

Model	Accuracy	Recall (danger)	This specific hard real clip
edge-fall-vlm-2.2b (SmolVLM2)	0.800	0.897	Missed
edge-fall-vlm-qwen3.5-2b (this model)	0.847	0.784	Detected

This is a genuine trade-off, not a strict improvement: SmolVLM2-2.2B has meaningfully better recall on the general in-the-wild population (catches more everyday falls), while this Qwen3.5-2B model has better overall accuracy/specificity and uniquely handled one particularly hard case (elderly subject, furniture occlusion, patterned-fabric camouflage, low light). Chosen here because quality/capability on hard real cases was prioritized over raw recall-rate optimization for this deployment — if maximizing general recall is the priority, edge-fall-vlm-2.2b may be the better choice; see that model's card.

Training data & recipe

Same as edge-fall-vlm-2.2b, for a clean architecture comparison:

Synthetic: 3D humans (Blender skin-mesh mannequin) rendered from many viewpoints and lens models incl. native fisheye, with exact labels; hard-negative normals (crouch/sit/kneel/bend).
Real negatives: UR Fall Detection dataset clips (research-use), oversampled, to fix specificity.
Full fine-tune (not LoRA), down3 label scheme, balanced oversampling.

Usage (transformers)

python

from transformers import AutoModelForImageTextToText, AutoProcessor
import torch
proc = AutoProcessor.from_pretrained("Luigi/edge-fall-vlm-qwen3.5-2b", do_image_splitting=False,
                                     size={"shortest_edge": 64*64, "longest_edge": 384*384})
model = AutoModelForImageTextToText.from_pretrained("Luigi/edge-fall-vlm-qwen3.5-2b",
                                                    dtype=torch.bfloat16).to("cuda").eval()
prompt = ("You are a safety monitor. These are consecutive video frames (oldest first), "
          "possibly with more than one person. Report whether ANYONE has fallen, fainted, "
          "is lying immobile, or is in distress; else normal. Answer with JSON only.")
frames = [...]  # list of PIL images, ~6 frames spanning ~3.5s
msgs = [{"role":"user","content":[{"type":"image"} for _ in frames]+[{"type":"text","text":prompt}]}]
text = proc.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)
batch = proc(text=[text], images=[frames], return_tensors="pt").to("cuda")
print(proc.batch_decode(model.generate(**batch, max_new_tokens=64, do_sample=False)[:, batch["input_ids"].shape[1]:], skip_special_tokens=True)[0])

Important: always pass enable_thinking=False to apply_chat_template — Qwen's chat template defaults to injecting a <think> block, which without this flag causes the model to ramble into open-ended reasoning instead of emitting the trained JSON format directly.

Limitations (read before use)

Research prototype, NOT a medical or safety-certified device.
Lower recall than the SmolVLM2 sibling model on the general in-the-wild benchmark — see the trade-off table above. Pick the model that matches your priority (general recall vs. hard-case robustness).
Still misses the fully-settled, heavily-occluded aftermath of the hard real clip that motivated this model — it catches the dynamic fall event, not every frame after.
Validated on small test sets (60-210 clips).
Not validated on elderly falls as a systematic category (the one hard clip is a single data point, not a validated subpopulation result).

License

Apache-2.0 (inherits Qwen3.5). Evaluation datasets (URFD, OmniFall/OOPS) are not redistributed; see their sources for terms.