Fall detection at the edge, privacy first.

1. Height decides whether the LIDAR can see a fallen person

A level scan plane above the floor passes over a person lying on it. Moving the sensor down to the floor fixed that for a person lying across or diagonal to the beam, but not end-on.

2. Floor trials with the legacy detector (floor-trials-1)

There were six lie-downs and a walking baseline. Lying across or diagonal to the beam was flagged (OBSERVE) within about a second of going down, at every range tried (1.3 to 2.6 m). Lying end-on, the plane saw only 0.22 to 0.33 m of soles or head, and both of those trials were missed.

3. The field recording exposed two escalation bugs. The fixes are tested offline and not yet enabled.

Every fix sits behind a configuration switch that defaults to the legacy behaviour. The replay harness runs the same detector code over the recorded bag on a laptop, without ROS.

4. The blind spot, seen by the cameras

Lying end-on with the feet toward the sensor, a person shows up on a floor-level LIDAR as two small clusters, and the detector raised 0 events. Stock RF-DETR, running on the Jetson, found the person on both cameras.

5. Stock RF-DETR on the LIDAR’s blind spot: conditions declared before the run

The evaluation plan, with its pass bars, was committed at 16:38 and the results at 17:26 (PDT, 2026-09-27). The run used 580 frames at 1 frame per second, with three model sizes served locally on the Jetson.

• The counter camera carries F. On the floor Brio, F was 26/30 (nano), 20/30 (base) and 30/30 (medium). B was 33/33 on both cameras.
• The walking boxes show hips and legs only. The counter camera cuts off heads, so C2 passes as a detection, not as a view of a standing person.
• C3 failed for every size. The median box width/height was 0.78 to 0.87 in lying segments B, D and E. The consequence was declared in advance: keypoints next. The first keypoint run was invalid under its own rule and was not scored. A rescore plan (v2) that fixes the class rule was written on 2026-09-28 and has not been run. A fine-tune that predicts the pose as a class is a third option (section 6).
• C4 is a median serial client round trip, with one request in flight and the cameras stopped. It is not a throughput benchmark. The lowest free memory (MemAvailable) was 2,584 MB. Whether inference disturbs the LIDAR under load was measured once on 2026-09-28, with a fine-tuned model: the scan stream held 10.009 Hz with no gap over 0.5 s while it served 580 frames (one run, cameras off; section 7).
• Near-duplicate frames. At 1 frame per second, each segment on each camera is closer to one trial than to thirty.
• On 09-27 the privacy settings were reported, not verified, and one of them did not work. The client posted only to 127.0.0.1:9001, but the server was also posting a record of every request to Roboflow once a minute (by the Inference 1.7.2 code; the capture cannot read TLS), and the opt-out setting in the container command (TELEMETRY_OPT_OUT) does nothing in Inference 1.7.2. It was found and switched off on 2026-09-28 (PDT), and the switch was verified in one captured run (one model, cameras off, no model download captured); see section 8.

6. A fine-tuned class split, on public data only

Box shape could not separate lying from standing (C3), so a third option for D1 was tested under a plan committed before any training started (2026-09-28, 11:30 PDT): fine-tune RF-DETR to predict the pose as a class. The data is public: three Roboflow Universe datasets, forked unchanged. The architecture search ran on one of them, the primary arm (arm A, 736 images), for 1 h 59 min. Arms B and C were plain RF-DETR nano comparison arms. Every model was trained and scored on Roboflow. No frame from the test room was uploaded, trained on or scored.

• Small and public. 73 test images, 89 instances (24 lying, 11 sitting), one dataset of one author’s rooms. Read the pass as not dead on arrival, not as a recall estimate.
• An upper bound. Nothing checked whether train and test share near-duplicate frames. Each arm has one scored training run, so seed variance is unmeasured (arm B ran twice by accident; the unscored duplicate moved test mAP@50 by 0.006).
• The errors are on the upright side. 4 standing instances were read as sitting and 1 was missed, with 0 false positives. lying stays 1.000 / 1.000 at every stored threshold from 0.11 to 0.87.
• Two evaluators. The model was picked on the search’s own valid score (96.85 mAP@50-95); every number above is Roboflow Model Evaluation’s (0.963 on the same split). Whether the rule would pick the same model under Model Evaluation is not known, because only two were evaluated.
• How the numbers were read. The 0.75 threshold comes from the valid split and was not tuned on test. Confusion matrices are stored at 0.1 steps only: F2 is read at 0.70 and 0.80, which agree, and the F4 swaps at 0.40. bed labels the furniture, so nothing here tests a person lying in a bed. Arm C’s splits were rebalanced to 70/20/10 before versioning, because its fork had 6 test images.
• Licences. The datasets are CC BY 4.0 as their Universe uploaders state (the URFD copy was not checked against the original dataset’s terms). The trained models are under Roboflow’s Platform Model License (PML-1.0).
• Not a room result. The room frames were scored under a separate plan (section 9). Source: the fine-tune plan and the results, section 3.

7. On the device: speed, memory, and the LIDAR under load

Later the same day (15:54 PDT), both fine-tuned models were pulled into the local Inference container on the Jetson and sent the 580 room frames of section 5, one request at a time, to 127.0.0.1:9001 only, with the cameras off. Their predictions were discarded by design: reading them would be a room result, which needs its own plan first.

• Against the stock models. The fast child is 27% faster than stock nano, and the rule’s pick costs what stock base and medium cost. The stock numbers are from 2026-09-27, measured the same way.
• Memory and the first call. The lowest free memory (MemAvailable) with both models resident was 2,799 MB. Only that minimum is usable, and with the cameras off the load of live capture is absent from it. After a restart, a model not yet in the cache pays a download and load on its first call: 17.5 s; with the weights already cached it was 1.8 s. That download has never been captured on the network.
• The LIDAR under load, CL1 and CL2. CL1, /scan count within 10% of idle: 594 against 567 (+4.8%), pass. CL2, no /scan gap over 0.5 s while the fast child served 580 frames: 0 (largest 0.105 s), pass. Both recordings read 10.009 Hz; the counts differ by the recorders’ spans. The bars were committed before the recording.
• One run of one model on one Jetson, cameras off. /fall_events stayed silent, and nobody was asked to be in the room. Nothing more is claimed about the detector under load.
• The speed figures are report only. F5 had no bar. It is a serial measurement with one request in flight, not a throughput benchmark. The LIDAR stack on the Jetson ran the branch fix/background-absorption, not main.
• Source. Fine-tune results, section 5.

8. What the on-device server sends: four network captures

Running inference on the device (DR-11) keeps frames in the room only if nothing else carries them out. On the afternoon and evening of 2026-09-28 (PDT; the repository dates the last three captures 2026-09-29 UTC), four tcpdump captures on the Jetson tested that, each against bars committed before it ran. TLS was not read: the destinations and sizes come from the captures, and what the bytes contain comes from the Inference 1.7.2 source code.

• The second bar of each pair passed. EG2, every non-LAN destination is a Roboflow host, passed in captures 1 and 2. VC2, the server answers within 120 s with the switch under test set, passed in captures 3 and 4: the first answer was seen 100 s after the server process started in capture 3 and 45 s after in capture 4. Both times are counted from that start, not from the container start, and in capture 3 the exact moment of the first answer was not captured.
• What capture 1 found. Until the post was switched off on the evening of 2026-09-28 (PDT), the server posted, once a minute, a record of every request, during every run of 2026-09-27 and 2026-09-28. It sent no frames and no boxes. The opt-out setting in the container command (TELEMETRY_OPT_OUT=True) does nothing in Inference 1.7.2. The bar was not moved. The capture saw 167 of the 580 frames, so 285,098 is a lower bound; that correction is written beside the result, not over it.
• Frames stayed on the device in capture 2. The full 166 MB of frames went to the local container and 13 KB left the device. One run, one model, cameras off. The model was already on the device, so no model download was captured.
• What still leaves. While the server answers requests, about 2.4 KB goes to api.roboflow.com every ~10 s: an aggregated usage record, with no image and no per-detection field. Its fields are listed under Known limitations. Whether that may leave a resident’s room was a product decision: yes, for now (Jeremy Gracey, DR-11). No revisit trigger is set.
• An idle container. A container that is not asked anything sent nothing to any non-LAN address for the 32.6 minutes it was watched. A model download has never been captured.
• Sources. DR-11, the fine-tune results, section 5, and the container command, jetson/inference-server-up.sh.

9. The fine-tuned model on room frames: one camera passed, one failed a pose

On 2026-09-29 the device-sized fine-tuned model was scored on the 580 recorded room frames of section 5, on the Jetson, at confidence 0.56. It is the faster of the two models in section 7, and not the one that passed F1 and F2 in section 6. The plan, the runner and the scorer were pushed to the public repository before the run. One run, scored once: 580 frames answered, no request error.

• One camera of two. The floor camera read the head-first lie-down as standing in 30 of 30 frames, and on its own it would fail R1. In the two diagonal segments it often returned no pose at all (20 of 35 frames and 11 of 33). Why was not examined.
• One subject, one room, one lie-down per segment. Frames within a segment are near-duplicates, so each count is closer to one trial than to thirty. Nobody fell. On the counter camera the walking frames show hips and legs only.
• Reviewed before the run. Two adversarial reviews of the plan and its scorer came before any result. The first found two ways a broken run could have been scored as a pass; both were closed before the plan was committed.
• Not verified at the run. The container’s settings were not inspected and the network was not captured, so what section 8 says still leaves is assumed to have left.
• D1 stays open. A passing bar is evidence, not a decision. Source: the room-frame plan and the results.