Privacy-Aware Customer Analytics in Physical Stores Using an Edge-Cloud Hybrid Artificial Intelligence System

Abstract

Customer analytics in physical stores must operate under three simultaneous constraints that are absent or weaker in online retail: limited on-site computation, strict personal-data privacy obligations, and the lack of a natural digital record of customer behavior. This study proposes an end-to-end edge-cloud hybrid artificial intelligence (AI) system that addresses these constraints jointly. On the edge side, an NVIDIA Jetson Nano device equipped with a Sony IMX477 camera performs motion detection, face detection, person detection, person re-identification, and demographic and affective attribute extraction (age, gender, ethnicity, emotion). Faces are detected and masked on the edge device, so unmasked facial biometrics are never transmitted to or stored in the cloud; person re-identification feature vectors are held only in volatile memory and purged daily. On the cloud side, scheduled batch tasks extract clothing, style, and color using Mask R-CNN, a VGG network, and k-means clustering with nearest-neighbor color mapping. Age, gender, and ethnicity are recognized by a single multi-task ResNet50 model trained on a merged UTKFace and FairFace dataset; emotion is recognized by a compact custom convolutional network trained on FER-2013. The system was validated end-to-end in a 45-participant in-store pilot study; per-class accuracy, precision, recall, and F1 values are reported with 95% confidence intervals given the pilot's small sample. The two core pipeline capabilities, person detection and person re-identification, reached 93% and 61% respectively under in-store conditions; the seven secondary demographic, affective, and fashion attributes (age, gender, ethnicity, emotion, clothing, color, and style) ranged from 57% to 88% accuracy. Face detection and masking succeeded on 100% of pipeline detections, so the privacy safeguard held under field conditions. During the pilot's 91.4-minute motion-gated monitoring session, chip temperature peaked at 51.5°C and memory briefly neared its 4 GB ceiling before settling well below it; sustained continuous-load behavior was not tested. The architecture’s principal contribution is that the edge-cloud split itself enforces data minimization, instead of relying on a compliance layer added after the fact. Limitations, including the small validation sample, illumination sensitivity, and domain mismatch in style detection, are reported transparently.