Random forest based classification of retroreflective foils for visible light sensing of an indoor moving object

Systems based on visible light sensing can relief
some of the anticipated challenges arising from the predicted
massive increase in connected Internet of Thing devices. For
example, identification and speed determination of mobile
objects can be achieved without the necessity to place actively
powered devices or sensors on the object itself. Instead, the
surfaces of the objects are simply equipped (coded) with
sequences of differently colored foils, which affect the respective
spectral compositions of reflected light. In this work, we present
an innovative approach for classifying differently colored
retroreflective foils in varying size configurations on a moving
object by utilizing the supervised machine learning algorithm of
random forest. For the respective experimental setup, consisting
of a single light source (as a transmitter) and a single RGB
sensitive photodiode (as a receiver for the reflected light from
the coded mobile object), we can show that not only the task of
identification, but also the task of determining the speed of the
object can be achieved with 98.8 % accuracy. By utilizing a
minimal feature set to create the random forest, the proposed
approach requires only minimal computational effort for model
generation and classification. The therewith-achieved results
are directly compared to an algorithm based on the more
complex and resource demanding method of Euclidian
distances. The satisfying congruence discloses the applicability
of the random forest model for such tasks, especially in scenarios
with highly limited memory resources and limited available
computational performance.