Inspiration
1 in 4 Americans over 65 falls every year. That's more than 14 million people. Falls are the leading cause of injury death for older adults, and they killed more than 41,000 people in 2023.
The fall itself is often not the worst part. In a BMJ study of adults over 90, 82% of falls happened while the person was alone, 80% of the people who fell couldn't get up on their own, and 30% lay on the floor for an hour or more. Doctors call this a long lie. It can lead to dehydration, pressure sores, muscle breakdown and pneumonia. In an older BMJ study, half of the people who lay on the floor for over an hour died within six months.
The usual answer is an alarm pendant, but it only helps if the person is wearing it, can reach it, and decides to press it. Most people in that study who had a long lie didn't use the alarm they had. We wanted a safety net that doesn't depend on the person who just fell. Most falls (79% of fall-related ER visits) happen at home, in the bedroom, the bathroom and on the stairs, so that's where SafeStep needs to work.
What it does
SafeStep is a camera-based fall detector that alerts a caregiver within seconds, with no button to press and nothing to wear.
- Detects the fall. A YOLO pose model tracks 17 body keypoints on each person in the frame. When the hips drop fast and the body ends up lying down, SafeStep flags a possible fall.
- Gives the person a chance to cancel. The caregiver gets a heads-up right away, and the person has a 30-second window to cancel. If SafeStep sees them stand back up, it cancels the alert automatically.
- Alerts the caregiver. If nobody cancels, the caregiver's phone gets an urgent push notification that says who fell, in which room and when, with I'm on my way, Call 911 and Map buttons.
- Escalates. If the person is still on the floor after 60 seconds, the alert is updated to say "still on the floor." If no caregiver responds, reminders repeat and a backup caregiver is alerted.
- Watches itself. The detector sends a heartbeat. If the camera crashes, freezes or gets unplugged, the caregiver gets a camera offline alert, so silence never looks like safety.
- Learns from patterns. Every event goes to a Databricks dashboard and a weekly email report. It shows which rooms and hours have the most falls, how quickly caregivers respond, and how often alerts were cancelled.
With Databricks Genie AI, caregivers can also explore the incident data by asking questions in everyday language, such as "Which room had the most fall detections?" or "How quickly did caregivers acknowledge alerts?" This gives them another way to investigate patterns alongside the dashboard charts, without writing SQL.
There's also a privacy mode that shows only stick-figure skeletons and never the video. That matters for cameras in bedrooms and bathrooms.
Impact: with SafeStep, a caregiver has an urgent alert within 30 seconds. With a median EMS response of about 7 minutes (13 minutes in rural areas), help can arrive in about 7.5–13.5 minutes instead of an hour or more on the floor. That's roughly 78–88% less time waiting.
How we built it
SafeStep has three parts that only talk to each other over HTTP, so each can fail or restart without taking down the others.
1. Edge detector (Python, Ultralytics YOLO pose, OpenCV, ByteTrack) It runs on a laptop or the Linux side of an Arduino UNO Q. For each tracked person, we measure body size in torso lengths, so the thresholds work at any distance from the camera. A fall needs a hip drop of at least 1 torso length within 1 second, then the torso staying more than 55° from vertical for 0.6 seconds. Each fall gets an evidence score:
$$ S = 0.40\,\hat{d} + 0.35\,\hat{v} + 0.25\,\hat{\theta} $$
Here $\hat{d}$ is how far the hips dropped, $\hat{v}$ is the peak drop speed, and $\hat{\theta}$ is the final torso angle. Each is scaled from 0 to 1 between a "weak" and a "strong" anchor value (for example, an angle of 45° counts as 0 and 85° counts as 1). The score is a heuristic for ranking falls, not a probability, and caregivers see it as low / medium / high / very high. The detector then follows the incident, sending immobility and recovered events linked to the original fall.
2. Alert service (Python standard library + ntfy) This is a small HTTP service that receives detector events, runs the cancel countdown and escalation, and sends push notifications through ntfy. The caregiver's "I'm on my way" button posts back through an ntfy topic, so the phone never needs to reach the home network directly (no port forwarding). It has no dependencies outside the standard library, so it runs on the UNO Q.
3. Analytics (Databricks) A shipper streams the alert service's event log into a Delta table with MERGE-based dedup and checkpoints that survive crashes. A SQL view rebuilds each incident from its events. On top of that are a Lakeview dashboard (12 queries) and a scheduled weekly email report.
Our Databricks workflow connects storage, analysis, AI, and reporting:
- Delta Lake: We use a Delta table as the storage foundation for our incident history in Databricks. It holds structured event records such as the room, camera, detection time, fall evidence score, immobility duration, and alert lifecycle status. The original JSON record is retained for additional details and future analysis. The analytics pipeline sends event metadata rather than camera video.
- Databricks SQL and SQL warehouse: Our Python shipper connects through the Databricks SQL Connector. SQL transformations reconstruct each incident from its lifecycle events, producing a shared incident view for analysis. A detection, notification, escalation, and acknowledgement can all belong to the same incident; grouping them correctly prevents follow-ups from inflating the incident count.
- Lakeview dashboard: The dashboard brings together room and time patterns, caregiver acknowledgement times, cancellations, escalations, notification delivery issues, and camera offline episodes. Its caregiver review page also shows reported outcomes and notes, helping distinguish confirmed falls, no-fall reports, uncertain cases, and incidents still awaiting review.
- Genie AI: Genie adds natural-language exploration of the SafeStep data alongside the dashboard. Caregivers can ask follow-up questions about rooms, response times, and incident history as they investigate a pattern. It makes the stored data easier to explore for people who do not work with SQL.
- Databricks Jobs and native dashboard emails: A scheduled Job refreshes the published dashboard on its SQL warehouse and emails a snapshot to configured recipients. Our default schedule is Sunday at 8:00 a.m. America/Chicago. The reporting job runs in Databricks even when the development laptop is off, and the snapshot follows the dashboard's configured date range.
This lets us examine both what happened and how the response worked. Repeated detections in one room could prompt a caregiver to check lighting or trip hazards, while slow acknowledgements or repeated escalations could prompt a review of caregiver coverage. These patterns guide follow-up; they do not by themselves establish why someone fell.
We also keep the metrics explicit: a cancellation is not automatically a false alarm, and a notification accepted by ntfy is not proof that a phone displayed it. Caregiver reviews provide a separate outcome signal. The shipper runs independently of the alert service, so a temporary Databricks outage does not delay urgent alerts; retained events can be ingested when connectivity returns.
We wrote tests for each layer, including end-to-end tests that send detector events through the real alert service.
Challenges we ran into
- Telling a fall apart from sitting down or lying in bed. Body angle alone flags someone lying on a couch, and speed alone flags someone dropping into a chair. We needed all three: a fast drop, then lying flat, and staying that way for a moment.
- Losing track of the person who fell. When someone hits the floor, the pose tracker often loses them and gives them a new ID, which would split one fall into two incidents. We added re-linking: a fallen person's track is kept for 30 seconds and matched to new tracks nearby.
- Avoiding alert fatigue. If caregivers get false alarms, they stop trusting the app. The cancel window, automatic cancel when the person gets up, and attaching follow-ups to the same incident all came from that worry.
- Failing loudly instead of quietly. A frozen camera looks exactly like a quiet house. We tied heartbeats to frames actually being processed, not just to the program running, so a frozen or unplugged camera triggers an alert.
- Running on limited hardware. PyTorch's default Linux build includes several GB of CUDA. We switched to CPU-only builds and kept the alert service dependency-free so it fits on an edge device.
Databricks limits. The SQL connector allows only 255 parameters per query, which capped our batches at 11 rows. We had to design idempotent, resumable ingestion around that.
Turning event logs into useful incident records. One fall generates several lifecycle events. In Databricks SQL, we had to reconstruct the incident without double-counting its follow-ups, and preserve the latest caregiver review even when later notification events arrived.
Accomplishments that we're proud of
- A complete pipeline from camera to caregiver: pose detection, fall detection, push alert, caregiver acknowledgement, and analytics.
- Getting help moving within 30 seconds without the person pressing a button, which addresses the main reason pendant alarms fail.
- Handling the hard cases, not just the demo: recovery, still on the floor, camera offline, unanswered alerts, and retries.
- Privacy mode, so a camera can go in the rooms where falls actually happen.
An analytics layer that turns alerts into prevention, by showing which rooms and hours are riskiest.
Connecting Delta Lake, Databricks SQL, the Lakeview dashboard, Genie AI, and scheduled Databricks reporting so caregivers can move from an urgent alert to a longer-term view of incident patterns.
What we learned
- The real problem is the wait. We started out thinking about detecting falls, but the research showed that the hour on the floor causes much of the harm. That shifted our focus to speed, escalation and reliability.
- A trustworthy heuristic beats a mysterious one. We kept the fall score explainable (drop, speed, angle) so caregivers and judges can see why an alert fired.
- Silence is a failure mode. Monitoring systems have to watch themselves.
How to design clear contracts between parts (a shared event schema) so our team could build the detector, alerts and analytics in parallel and connect them at the end.
Analytics depends on clear definitions. Counting detected falls, caregiver-confirmed falls, cancellations, and notification deliveries as separate measures gives caregivers a more accurate picture. Databricks helped us bring these signals together while keeping their meanings visible.
What's next for SafeStep
- Real-world validation on public fall datasets and with care facilities, to tune thresholds and measure detection and false-alarm rates.
- Multi-camera coverage for whole homes and facility wings, with a staff dashboard for nursing homes and hospitals.
- Fall-risk prediction, using gait and near-falls to warn caregivers before a fall happens.
- Two-way voice, so the person can say "I'm okay" or "Help" without touching anything.
- On-device inference on lower-cost hardware, so a SafeStep camera can be cheaper than a monthly pendant subscription.