feature learning model and a location estimation function.
Secondly, the location and the orientation together can be
sent to the retailer server, so as to rank the possible retailers
according to the mall layout. We developed a retailer ranking
algorithm, based on the overlap degree between the phone’s
angle of view and the retailers in the view. Once the
retailers are ranked, their corresponding reviews can be
queried through the augmented reality engine. The retailer
reviews are crawled periodically by the iterative learning
to query model and saved into the retailer DB. Finally, for
smooth content display in augmented reality, the client app
keeps detecting changes w.r.t. the phone location and its
orientation. If there is any change, then augmented reality
engine will be called to update the localization and retailer
ranking results for display.
We deployed IntelligShop in a real mall of Singapore (see
our demo listed in Footnote 1). Take the first floor as an
example. The public space outside the retailers is 45×36 m2.
It has totally six retailers, and it is equipped with WiFi. For
localization, we sampled 16 locations and used a Samsung
S3 phone to collect data. The localization model was built
and tested with a Samsung S4 phone. As we show in the
demo, we can improve the localization accuracy9 by 34%
given this device heterogeneity.
It is worth noting that, before enabling the IntelligShop
app in a mall, we need to survey the mall. This survey
consists of: i) collecting WiFi signal data and the mall’s
floor plan to build the indoor localization model; ii) gathering
retailer names (and their branch names if any, for
disambiguation) to crawl the reviews. Such surveying can
be eased with the mall layout and tenant information, which
is provided by the mall owner or obtained from the publicly
available online sources.