CamGuard

Inspiration

CamGuard was inspired by a simple but painful reality: falls from beds are one of the most common preventable injuries for elderly patients and infants. Caregivers cannot monitor someone every second, and traditional baby monitors or hospital cameras only record footage — they do not understand risk.

We wanted to build something proactive instead of reactive. Not just a camera that watches, but an AI agent that predicts and intervenes before a fall happens.

What CamGuard Does

CamGuard is an agentic AI bedside safety system that predicts and prevents falls in real time.

It detects early fall-risk signals such as:

Edge drifting toward the mattress boundary
Unassisted sit-ups
Leg swing off the bed
Attempted bed exits
Progressive restlessness trending toward unsafe positions

Instead of triggering on every movement, CamGuard models dynamic risk over time:

$$ R(t) = w_1 E(t) + w_2 S(t) + w_3 L(t) + w_4 M(t) $$

Where:

(E(t)) = edge proximity score
(S(t)) = sit-up posture probability
(L(t)) = leg-off-bed likelihood
(M(t)) = motion instability metric
(w_i) = adaptive weights personalized per user

If risk exceeds a threshold ( \theta ), the agent decides whether to:

Send a caregiver notification
Escalate if unacknowledged
Log the outcome and adapt thresholds

This makes CamGuard adaptive, not static.

How We Built It

1. Real-Time Pose Estimation

We implemented body keypoint detection to track hips, knees, shoulders, and torso orientation. This allows us to classify transitions such as:

Lying → Sitting
Sitting → Edge shift
Sitting → Standing attempt

2. Virtual Bed Boundary Modeling

We mapped a safety boundary over the mattress and computed normalized edge distance:

$$ D = \frac{x_{\text{body}} - x_{\text{center}}}{\text{bed width}/2} $$

As (D \to 1), fall probability increases significantly.

3. Agentic Decision Layer

Rather than hard-coded rules, we implemented:

Temporal smoothing of risk signals
Adaptive threshold tuning
Feedback-driven weight updates

Each incident outcome updates the model nightly to reduce false alarms.

4. Escalation Logic

We built a tiered alert system:

Level 1 — Low confidence warning
Level 2 — High confidence fall risk
Level 3 — Escalation if no caregiver response

Challenges We Faced

False Positives

Babies roll frequently. Elderly patients reposition often. Early versions triggered too many alerts.

We shifted from binary detection to risk accumulation over time to smooth noise.

Personalization

Movement patterns vary significantly between individuals.

We introduced adaptive weight updates to learn per-user risk profiles.

Real-Time Constraints

The system had to respond instantly. We optimized inference latency and reduced computational overhead for edge deployment.

Context Awareness

Not every sit-up is dangerous. Context matters.

We incorporated environmental and temporal modifiers into the risk calculation.

What We Learned

Proactive AI requires temporal modeling, not just classification.
False positives destroy trust in safety systems.
Agentic systems must include feedback loops.
Personalization dramatically improves reliability.

Future Work

On-device privacy-preserving inference
Multi-camera depth fusion
Reinforcement learning for smarter escalation policies
Integration into hospital bed systems

Conclusion

CamGuard is not just a monitor — it is a learning bedside safety agent designed to prevent injuries before they happen and protect the most vulnerable.