Predicting consumers' locations in dynamic environments via 3D sensor-based tracking

Brick-and-mortar stores for non-food items allow customers to quickly try items and take them home, but lack certain convenient features of online shopping, such as personalised offers and recommendations for items of possible interest. Mobile in-store shopping applications would allow to combine these advantages if they would derive current customer needs from customer activities. A natural way to infer interests of shops' visitors is to analyse their motion and places where they stop. This paper presents a low-cost depth sensor -- based people tracking system and a method to predict future customer locations, developed for environments where items are frequently re-located and customer routes change accordingly. The tracking system employs adaptive background modelling approach, allowing to quickly distinguish between moving humans and re-located objects. Similarity-based location predictions use fairly small datasets of recent tracks of other customers and allow predicting locations of future stops very soon after monitoring starts: after just a few minutes. Therefore this approach also provides for imperfect tracking, e.g., due to occlusions. In the tests with the data, acquired in a real clothing and cosmetics department during 50 days, future places of customers' interest were predicted with an average accuracy 60% and an average distance error half a metre.