Aerospace and Electronic Systems Magazine July 2017 - 53

Li et al.

Figure 2.

Block diagram of the design flow in LDspectral.

CASE STUDY: BACKGROUND SUBTRACTION
We demonstrate the importance of BSS, which is a core aspect of
the LDspectral design methodology discussed previously, through
a case study involving background subtraction. The two metrics
that we consider in this evaluation are the accuracy Fmeasure of the
background subtraction (foreground extraction) results, and the
average execution time tave to extract the foreground. We focus
on quantifying trade-offs consisting of singleton (one-band) and
two-band subsets, and demonstrate significant variations in performance trade-offs among different subsets. Analysis and optimization of BSS trade-offs among larger subsets (i.e., where the subset
size exceeds 2) are motivated through this preliminary study.
The band subset processing subsystem for this multispectral
background subtraction case study is illustrated in Figure 3. In the
context of this case study, the illustration represents the internal
functionality associated with the block in Figure 2 that is labeled
band subset processing. In the dataflow graph subsystem depicted
in Figure 3, each actor reads a pointer to an image from its input
buffer, and outputs a pointer to the image that results from the image processing operation performed by the actor.
We use LIDE to develop a prototype implementation of the
band subset processing subsystem in Figure 3, and apply calls to
selected OpenCV functions in some of the actors within this implementation. The image read actor in Figure 3 is used to inject a
stream of pointers to successive images into the subsystem so that
background subtraction can be performed separately on each image that is referenced (pointed to) in the stream. At the output of
the image read actor, each image contains a set of m separate comJULY 2017

Figure 3.

Block diagram of band subset processing in the background subtraction
system.

ponents, where each component corresponds to one of the selected
spectral bands (i.e., an element of the set S, as defined previously).
The image combination actor then performs pixel-level fusion to
combine the components associated with the selected bands into a
singled "fused" image.
The background subtraction actor then computes an initial
background subtraction result and passes the extracted foreground
through the image pointer produced on its output. The core background subtraction operation applied by this actor is carried out by
the OpenCV function called BackgroundSubtractorMOG2, which
applies a Gaussian mixture model (GMM) [21], [22].
The foreground filter actor in Figure 3 is designed to remove
noise from the output of the background subtraction actor. In the
foreground filter actor, we use two morphological operations-
erosion and dilation-through their respective implementations
in OpenCV. Intuitively, the erosion function helps to remove

IEEE A&E SYSTEMS MAGAZINE

Table of Contents for the Digital Edition of Aerospace and Electronic Systems Magazine July 2017