RFC: Improvements to drivers/sensors for light sensors

[microbuilder]

RFC: Improvements to drivers/sensors for light sensors

Introduction

This RFC details a number of issues and potential improvements to the handling
of 'light' in the current sensor API.

An explanation of radiometric versus photometric units is presented,
as well as some options to make the current system more practical in
real-world applications.

Problem Description

The current solution for 'describing' light-related sensor data is
fundamentally flawed at the unit level (lux for IR, etc.), and has some
unnecessary limitations in terms of what type(s) of light-related
data can be stored or conveyed with the current API.

Proposed Changes

• The current value types assigned to existing light channels should be
  changed from erroneous photometric types (lux) to an appropriate radiometric
  value type.
• Support for multiple instances of a channel -- with appropriate minimal
  meta-data -- should be added to support classes of sensors that
  respond with a range of data samples in a single fetch event.
• A wavelength-neutral light channel should be introduced for
  less rigidity when working with non RGB or IR light sensors.
• A new dedicated 'lux' channel should be added to correctly accommodate
  the measurement of human vision as a scientifically important photometric
  value.

Detailed RFC

At present, sensor.h has the following photometric value types defined
for light sensors:

/** Illuminance in visible spectrum, in lux. */
SENSOR_CHAN_LIGHT,
/** Illuminance in infra-red spectrum, in lux. */
SENSOR_CHAN_IR,
/** Illuminance in red spectrum, in lux. */
SENSOR_CHAN_RED,
/** Illuminance in green spectrum, in lux. */
SENSOR_CHAN_GREEN,
/** Illuminance in blue spectrum, in lux. */
SENSOR_CHAN_BLUE,

There are a number of problems and limitations with the current approach, which
are detailed below.

Key Technical Concepts

Before detailing the problems with the current sensor API, an explanation of
radiometric versus photometric units may be useful.

Radiometric vs. Photometric Units

Radiometric Units

Electromagentic radiation is characterised by radiometric units, which
describe the physical properties of light (the number of photons, photon
energy, or radiant flux). These units have no relation to human vision.
Infrared radiation, for example, is not visible to the human eye (>780 nm) but
it definitely exists as a radiometric phenomenon and can be accurately
measured, described and analysed for scientifically significant purposes.

Photometric Units

To characterise light relative to the human eye, we need to use
photometric units such as luminous intensity, which represents the
light intensity of a source as perceived by the human eye, measured in
candela (cd).

Since photometric measurements are limited to what the human eye can
perceive, these measurements are restricted to the visible spectrum
of 380 nm to 780 nm wavelengths.

One of the most common photometric units is luminous flux, which is
measured in lumens (lm). It's even more common derivative is illuminance
(used in this RFC and the current API), which is the luminous flux incident per
a specific area. Illuminance is measured in lux (which is equal to
lm/m^2).

It should be noted that all photometric units are based on a theoretical
model of human vision, and thus are not entirely accurate (compared to
radiometric units). Anything dealing with 'lux' or photometric units will
always, by definition, be an approximation, even if it's a reasonably
good one!

Photometric Units and Radiometric Equivalents

The two unit families have the following relationships:

Photometric Unit	Dimension	Radiometric Unit	Dimension
Luminous Flux	lm	Radiant Flux	W
Luminous Intensity	lm/sr=cd	Radiant Intensity	W/sr
Illuminance	lm/m^2=lux	Irradiance	W/m^2
Luminance	cd/m^2	Radiance	W/(sr m^2)

Key Problems

Some of the key limitations of the current approach to capturing light samples
are listed below:

1. Lux is a photometric unit. Photometric units are used to represent
   light as perceived by the human eye. Radiometric units would be
   preferable since they are based on an actual physical model, not a
   theoretical estimation, and you can easily convert from radiometric to
   photometric units. (The radiometric equivalent of illuminance or lux is
   irradiance, which is measured in W/m^2.)
2. Lux, by definition, has no meaning in the IR, RED, GREEN, or BLUE
   spectra since it can't be isolated to these narrow wavelengths, but is
   a calculation based on a curve representing a model of 'normal' human
   vision. Lux should be it's own inferred value type, based on raw
   radiometric data capture and the response of the photodiodes used.
3. IR, Red, Green and Blue are common components of the light spectrum, but
   there is currently no way to account for other (visible or non-visible)
   wavelengths which are scientifically important, such as UV, or narrow
   bandwidths that can be captured by speciality light sensors or photodiodes.
   Ideally, it should be possible to return a radiometric sample that indicates
   the wavelength(s) of the light sample, and it's relative strength, without
   an arbitrary limitation to the IR/R/G/B ranges.

Real World Example: Spectrometer

An easy to understand example of the limitations of the current API is
trying to design a spectrometer.

A spectrometer works by capturing a number of samples at fixed wavelengths or
wavelength ranges, with all of the spectral data in that range making up one
dataset (capture in a single 'fetch' event).

A spectrometer might measure light in the human-visible-range, for example
(380-780nm), and in 10nm intervals, meaning that each 'fetch' operation
consists of 40 samples, each associated with either a fixed wavelength
(ex. 420 nm, 430 nm, etc.) or a range of values (385-395 nm).

At present, there is no obvious mechanism to support either of these
requirements:

• Multiple instances of a channel
• Appropriate meta-data per channel instance

Proposed Changes/Solutions

1. Converting between radiometric and photometric units is done via the
   luminous efficiency function, of which several versions exist (CIE 1931,
   CIE 1988, etc.). Since lux is a commonly used value type, conversion
   from irradiance to illuminance should be provided via an appropriate
   function, taking radiometric data as an input and lux as an output.
   Radiometric units should be preferred, with support added for converting
   between the two unit families.
2. A new SENSOR_CHAN_LUX channel should be added, or the current
   ambiguously-named SENSOR_CHAN_LIGHT channel should be renamed to clearly
   identify this as a photometric unit (lux) describing an estimation of
   human vision related to the capture light sample(s).
3. Changes should be made to enable multiple instances of a channel per fetch
   operation, and the addition of meta-data to distinguish each of these
   channel instances.

Concerns

These modifications would obviously be a significant breaking change to the
current sensor API, and require changes to any drivers using the current API.

Memory usage is also a critical concern, particularly when you need to maintain
a large number of data samples in memory.

Unresolved Questions

• How can we best support multiple instances of a channel per fetch event,
  without unnecessarily increasing memory usage per data sample or complicating
  the sensor API?
• How can we minimize breaking changes to the existing API to avoid having to
  modify all existing sensor drivers and related application code?
• Is a mechanism for returning 'raw' sensor data (lower memory requirements)
  useful?

Related Issues

#1387: Lack of a performant sensor API

[microbuilder]

An initial attempt to address some of the weaknesses and errors in the current light sensor API is visible in PR #14245

[adamfowleruk]

I'm interested in this given the need to support the detection of being indoor or outdoor for a digital contact tracing application. Current sensor we are looking at are in this family: Broadcom APDS-9250 - Has components for IR,R,G,B.

[teburd]

Proposal for better sensor channel specifiers #63830 would perhaps atleast address the issue of the type of data (LIGHT) subtype (R/G/B/IR/UV/CUSTOM) and channel index in the sensor (N) to deal with light spectrometer sensors for example.