14.5Retail and Smart Spaces

Retail and smart-space intelligence reduces to one persistent spatial problem: jointly associating, across a network of calibrated overlapping cameras and over time, which person is interacting with which product at which shelf location. Unlike the warehouse logistics of Section 9.5, where the environment is engineered for the robot, here the space is a crowded store full of untracked people, and its distinguishing challenge is dense multi-object tracking with reliable human-object association through frequent mutual occlusion. Figure 14.5 shows how the major applications share this core. Multi-camera tracking provides the spatial backbone, with benchmarks such as WILDTRACK [2999] and MMPTRACK [3000] and methods that project per-camera features onto a shared ground plane to detect and track people across views [3001]; this is the machinery behind cashierless checkout, where a shopper’s path and pickups must be reconstructed without a checkout line. Shelf monitoring turns perception toward the products, detecting densely packed items for inventory and planogram compliance, supported by the dense-detection benchmark SKU-110K [3002] and by methods that formulate compliance as matching the observed shelf layout against the ideal one as a sub-graph isomorphism [3003]. In-store analytics aggregates the same tracks into shopper-flow heatmaps and layout insights, building on cross-camera re-identification [3004] to stitch one customer’s trajectory across non-overlapping views. The deployed layer includes cashierless systems such as Amazon’s Just Walk Out¹ and platforms from Trigo,² AiFi,³ and Zippin,⁴ alongside shelf-analytics vendors such as Trax.⁵