Question 1

How does the simulation compute a worker's pose from a 3D skeleton?

Accepted Answer

Each worker is a real THREE.Group joint hierarchy: pelvis, spine, chest, neck and head form the torso chain, while hips, knees and ankles form each leg and shoulders, elbows and hands form each arm. Every joint is a THREE.Object3D whose rotation and, for the pelvis, vertical position are driven procedurally to produce walking, squatting, bending and falling motions. Every frame the simulation reads each joint's real world-space position (via getWorldPosition on the live matrix hierarchy) and computes genuine joint-angle features -- spine tilt from vertical, hip and knee flexion, limb speed -- using real vector dot products and angle-between-vector formulas, exactly as a video-based pose-estimation pipeline would work from detected keypoints.

Question 2

What is the anomaly score and how is the safe-motion baseline built?

Accepted Answer

At startup the simulation builds a hidden calibration skeleton and drives it through the same procedural pose function used for the visible workers, but restricted to safe walking and safe squat-lifting motions, sampling roughly 150 poses across the gait and lift cycles. From those samples it computes a per-feature mean and variance -- a real baseline distribution over the feature vector [spine angle, hip flexion left/right, knee flexion left/right, limb speed, torso angular velocity]. Every live frame, each worker's current feature vector is converted into a per-feature z-score against that baseline and combined into a single RMS z-score, equivalent to a Mahalanobis distance under a diagonal covariance assumption. A pose that matches the safe baseline scores near zero; a fall or an extreme spinal bend scores far higher because several features deviate from their safe-motion mean simultaneously.

Question 3

How does the simulation detect an unsafe lifting angle specifically?

Accepted Answer

Manual-handling guidance (e.g. HSE and NIOSH lifting equations) treats lifting with a bent back and straight legs as high-risk, versus squatting with bent knees and a near-vertical spine. The simulation checks this directly from geometry: it flags an unsafe lift when the spine-to-vertical angle exceeds roughly 38 degrees while the average knee-flexion angle stays below about 16 degrees (legs still nearly straight) during a non-walking action. This mirrors the real biomechanical distinction between a safe squat lift and a dangerous stooped lift, computed from the same joint vectors that drive the anomaly score.

Question 4

How is a fall distinguished from an anomalous but non-fall pose?

Accepted Answer

Falls are detected with an explicit geometric rule in addition to the statistical anomaly score: if a worker's pelvis joint drops below about 55% of their standing hip height, the pose is flagged as a fall regardless of the current z-score. This catches falls immediately and reliably, since collapsing to the ground is the most extreme and unambiguous departure from the safe-motion baseline -- it also drives the RMS z-score very high on its own, because spine angle, pelvis height and limb speed all move far outside their baseline ranges at once.

Question 5

How does restricted-zone detection work in 3D?

Accepted Answer

The restricted zone is a real axis-aligned rectangular region defined in world (x, z) floor coordinates and rendered as a translucent red overlay when enabled. Every frame, each worker's pelvis world position is tested against the zone's min/max bounds on both axes -- a straightforward point-in-region containment check performed on live 3D coordinates, the same approach real site-monitoring systems use with geofenced zones derived from camera calibration or floor-plan coordinates.

Question 6

Why use articulated procedural motion instead of pre-recorded motion-capture clips?

Accepted Answer

Procedural motion lets every joint angle, and therefore every feature the anomaly detector reads, be generated from an explicit, inspectable formula rather than baked into opaque animation data. That keeps the whole pipeline auditable end to end: the same pose function drives both the visible skeleton and the baseline calibration sampling, so the safe-motion envelope is guaranteed to be geometrically consistent with what the workers actually do on screen, and triggered fall or unsafe-lift events are genuine departures from that same function rather than scripted visual effects.

Question 7

How does this relate to real construction-site AI safety monitoring?

Accepted Answer

Commercial site-safety platforms (e.g. Newmetrix, Human Condition Safety, Intenseye) run 2D pose-estimation models such as OpenPose or MediaPipe over CCTV feeds to extract skeletal keypoints per frame, then apply rule-based or learned anomaly detectors over derived joint angles and velocities to flag falls, unsafe lifting posture, missing PPE and geofenced zone breaches, alerting site supervisors in real time. This simulation reproduces the same conceptual pipeline -- skeletal keypoints, geometric feature extraction, baseline-relative anomaly scoring, zone geofencing -- with a synthetic, fully client-side 3D skeleton instead of a live camera feed and a trained neural network, so it should be read as a mechanism demonstration rather than a production safety system.

🦺 Bouwplaats-veiligheidsmonitor — Pose-anomalie Live

Over de Bouwplaats-veiligheidsmonitor

Veelgestelde vragen

What did you find?