The noise floor is the background power level when no signal is present. Many detectors first estimate it, then judge a signal by how far it rises above it. A typical example is finding WiFi bursts in a band that also holds receiver noise and other weak transmitters. It builds on Thermal Noise and Noise Figure.

Why Estimate It

  • every receiver, gain setting and place has a different floor
  • the floor is not flat across frequency, because of filter roll-off and interference
  • a fixed threshold fails as soon as the floor moves

Percentile Method On A Spectrogram

Steps:

  • for each frequency bin, take a low percentile (for example the 5th) of the power over time
  • signals are usually on for only part of the time, so the lowest values in a bin are background
  • smooth the result across frequency (for example a 61-bin moving average) to remove jitter between bins
  • subtract it and clamp at zero, which gives a flat background of zero

Where it fails:

  • a signal that is on almost all the time in a bin (more than 95% of frames for the 5th percentile) raises the estimate, and part of the signal is subtracted

Robust Spread With MAD

  • the median absolute deviation measures the spread of the noise
  • a few strong signal values hardly move a median, but they move a standard deviation a lot
  • 1.4826 makes the result equal to when the noise is Gaussian
  • useful for generating extra noise at the right level, for example to fill new bandwidth when a capture is upsampled

Fast Estimate From The Lowest Values

  • take every 10th pixel of the spectrogram
  • find the lowest 30% with a partial sort (std::nth_element in C++, average )
  • the floor estimate is the mean of those values
  • fast enough to run on every frame in a live system

From Floor To Detection

  • set the threshold a fixed margin above the floor, so the false alarm rate stays about the same when the floor moves
  • this is the idea behind CFAR (constant false alarm rate) detectors in radar