Looks like there's a few good ideas in there.

  My summary:
1) Your space is anisomorphic; if I pick the word correctly; that is, it
is not of uniform texture.  Almost any scanning modality will of
necessity detect irregularities.

2) Probably the only useful modality to distinguish important objects
(humans in particular) by characteristic is heat, and that only useful
at night.

3) The alternative is image comparison.  The "reference" image may be
either fixed or updated by some delay (low-pass filter) mechanism.

4) A mixture of the above would seem most appropriate.  PIR detectors,
as already installed, are the simplest implementation of this, but give
a view with no spatial dimension (ie., only time!).

5) What is therefore called for is a two or possibly three-dimensional
implementation.  The simplest way to do this to my mind is an a small
array of linear (one- dimensional) resolvers strategically placed.  Such
devices are readily constructed using PIR sensors and drum scanners,
both readily available in the "disposals" (surplus) market and of
course, new.

 6) Processing of the data stream from these is probably best performed
centrally.  Each sensor requires a "reference" store, some processing to
separate differential data and (low-pass) update the reference, and a
mapping function of its linear "view" into (part of) the actual area in
conjunction with overlapping sensors.  It would be appropriate as
suggested, to implement a "teaching" function of the fully implemented
system to map critical points and subsequently interpolate between
these.  The map would become an array of point entries, each documenting
an actual grid location, an identity/ position (one-dimensional) pair
for each sensor which can "see" this location, and an az/ el pair to
direct the camera to this location.

  The algorithm to interpolate (map) readings is most challenging.  It
goes along the lines of: If one sensor only registers, approximate by
averaging (least squares) those reference points with the closest
readings, giving preference to points ONLY seen by that sensor.  If two
or more sensors register, use points with entries for those sensors
accordingly.

  Not trivial.  It is either formal or informal "fuzzy" logic!

  Cheers,
        Paul B.