AI Accuracy Manual

For the system to detect objects reliably, the camera optics and installation are essential. These general rules should be followed to optimise the field of view and the accuracy of the AI engine:

• • Mounting Height:

• A standard height of 3 metres above ground level is recommended to balance the viewing angle and coverage.

• • Tilt Angle:

• The angle between the lens and the horizontal should be between 0° and 45°. Exceeding this angle increases blind spots and distorts the workers’ silhouette, reducing the effectiveness of the pose model.

• • Occlusions and Line of Sight:

• The system requires a clear, direct line of sight to the target.

  • · Occlusions (objects that block the view such as columns, heavy machinery or shelving) can prevent the AI from detecting a person or their protective equipment.

  • · If a worker is partially hidden, the model’s confidence drops; cameras should therefore be positioned to minimise physical obstacles in the highest-risk areas.

• • Environmental Conditions:

• The scene background should not be overly complex or contain constantly moving elements other than the detection target. Adequate lighting and high contrast between PPE (such as high-visibility vests) and the background will drastically improve the confidence rate.

Given the diversity of supported manufacturers, we recommend and use camera simulators integrated with the installation layouts. This pre-production phase makes it possible to:

• • Verify Pixel Density:

• Validate whether the required thresholds (e.g. 125 px/m for recognition) are met in critical areas.

• • Optimise the Field of View (FoV):

• Avoid blind spots through virtual positioning on the floor plan.

• • Optics Selection:

• Confirm whether the chosen camera and lens model is suitable for the required detection distance before physical installation.

Detecting personal protective equipment (helmets, vests, gloves) requires a specific information density so that the YOLO engine can distinguish items with high confidence. Below we describe some pixel densities by equipment type, as examples:

• • Helmets:

• Require a size of 30–50 píxeles.

• • Reflective Vests:

• Being more visible, they are detected with high accuracy from 50–100 píxeles. High-visibility colours (fluorescent yellow/orange) dramatically improve the model’s detection rate.

• • Gloves and Glasses:

• Due to their small size, the system requires 20–30 píxeles. The use of longer focal length optics (zoom) is recommended for these applications. Uniform lighting that avoids harsh shadows on the equipment is essential.

For technical enquiries about hardware compatibility or custom quotes, contact us at:

• • Technical Support: support@safe.ai

• • Sales: commercial@safe.ai