Image requirements for the Human pose detector and its sub-tools

Hardware requirements for neural analytics operation

Requirements for the operation of the Human pose detector and its sub-tools are given in the table:

Camera
  • Resolution must be at least 640x360 pixels. We also don't recommend using a resolution higher than 1920×1080, since higher resolution doesn't increase the detection quality, but significantly increases the consumption of resources. The optimal resolution for solving typical tasks is 1280x720
  • The frame rate per second in the video stream from the camera is at least 8
  • Color image only
  • Correct operation of the detection tool isn't guaranteed if you use frames that were masked before they got into Axxon One by an IP camera or similar methods.
  • Correct operation of the Human pose detector tools isn't guaranteed if you use fisheye cameras and other cameras with strong image distortion.
  • The camera tilt angle from top to bottom on a person must not exceed 30°. Otherwise, the detection tool won't operate correctly.
    The following diagrams show the examples of:
    • The acceptable tilt angle:
    • The unacceptable tilt angle:


Lighting
  • Lighting in the scene is at least 50 lux per square meter. In conditions of insufficient or excessive lighting (night or light-striking), stable operation of the analytics isn't guaranteed
  • No abrupt changes in lighting
Scene and camera angle
  • Moving objects are visually separable from each other
  • The background is mostly static and doesn't change abruptly
  • Moving objects are minimally obscured by static objects in the scene (columns, trees, and so on)
  • The analyzed scene doesn't have reflective surfaces and sharp shadows from moving objects. If present, they must be masked
  • Camera shake doesn't result in image offsets greater than 1% of the frame size
  • For the correct operation of the Sitting Person Detection and Person Down Detection sub-tools, the area of interest must be covered from at least two perpendicular angles (a and b are the directional beams from the cameras and the point of their intersection is the area of interest):

    The algorithms have a limitation and won't work correctly if:
    • a person falls down or sits down with their head or face towards the camera,
    • a person falls down with their feet towards the camera (for the Person Down Detection sub-tool)