Edge-Vision for Smart Elevators: Privacy-Preserving Headcount Analysis

Modern smart‑building infrastructures increasingly rely on distributed edge‑vision sensors to monitor elevator traffic for operational efficiency, maintenance planning, and accessibility compliance. However, the continuous video stream raises significant privacy concerns, especially in residential and mixed‑use towers where passengers expect anonymity. In this work we present **Edge Vision Smart Elevators (EVSE)**, a framework that performs accurate, real‑time headcount estimation directly on commodity edge AI accelerators while guaranteeing differential‑privacy (DP) protection for every captured frame. EVSE combines a lightweight pose‑free person detector based on MobileNet‑V2 with a temporal‑aggregation module that fuses short‑term tracks into a private headcount via the Gaussian mechanism. The system operates at $\le$ 15 W on an NVIDIA Jetson Orin module, achieves a mean absolute error (MAE) of 0.21 persons per floor‑crossing ($\approx$ 4 % relative error on a 15‑story testbed), and satisfies ($\varepsilon$ = 0.5, $\delta$ = 10⁻⁵)‑DP guarantees with a utility loss of < 2 % compared to the non‑private baseline. Ablation studies show that the privacy module contributes only 1.3 % of the total inference latency, confirming that strong privacy can be attained without sacrificing real‑time performance. We discuss deployment considerations, trust‑worthy data handling, and outline extensions to multi‑camera floor‑supervision and encrypted‑log aggregation for building‑wide analytics.