Suning’s ‘Beidou’ AI System Transforms Retail Customer Flow Analytics

Introduction

With the deepening of the Internet+ era, offline retail is undergoing disruptive change and reshuffling. Suning, a pioneer in smart retail, has seized this opportunity to launch the Beidou system, a comprehensive AI‑powered platform for store analytics.

Purpose of the Beidou System

Beidou aims to provide precise, fine‑grained customer analysis for physical stores, turning raw video and WiFi data into actionable insights for marketing, layout optimization, and operational decision‑making.

Key Metrics for Physical Stores

Critical performance indicators include foot traffic before the store, in‑store traffic, transaction count, sales amount, and average transaction value. Accurate foot‑traffic data underpins strategy, investment decisions, and performance evaluation.

Evolution of Counting Methods

Human Counting

Early stores assigned staff to manually tally visitors, which suffered from fatigue, missed counts, and high labor costs.

Staff attention wanes over time, leading to missed counts.

Labor costs far exceed equipment investment, especially for large stores.

WiFi Counting

With widespread free WiFi, stores began counting customers via MAC addresses and IP tracking. However, this method faces three major drawbacks:

WiFi signal must fully cover the monitoring area and remain stable.

Customers must connect to WiFi for their data to be captured.

Signal offset (6‑10 m) can cause significant data distortion.

Video Counting

Advances in AI and computer‑vision enable video‑based counting. By processing existing surveillance streams, the system extracts, recognizes, and tracks moving objects to generate complete, real‑time customer flow data.

The video counting pipeline includes three key stages: face detection, optimal face capture, and face‑recognition matching.

Face Detection

High‑definition cameras capture video streams, allowing real‑time detection of up to 20 faces per frame, robust to pose, angle, and partial occlusion thanks to deep‑learning models.

Optimal Face Capture

From 30 fps video, the system selects the highest‑quality face image using orientation analysis, blur detection, and a quality‑scoring model.

Face‑Recognition Matching

Using a refined SoftmaxWithLoss formulation (inner‑product bias removed, L2‑norm on weights and inputs) and an angle‑margin loss, the model achieves 99.70 % accuracy on the LFW benchmark, close to the state‑of‑the‑art 99.83 %.

To further improve large‑scale one‑to‑many recognition, the team added metric‑learning fine‑tuning with triplet loss, boosting accuracy by roughly seven percentage points.

System Architecture

Beidou integrates background modeling, facial recognition, 3D depth information, multi‑target tracking, and deep‑learning networks into a unified solution. The tracking stack combines KCF for initial target classification and Kalman filtering for real‑time multi‑target tracking, achieving the required speed and accuracy.

The detection backbone builds on a lightweight SSD network, further optimized for Suning’s specific scenarios, and runs on a GPU cluster of six high‑performance servers, reducing model update cycles from weeks to hours.

Intelligent Services Enabled by Beidou

Smart Guidance: Recognize returning customers and provide personalized recommendations; guide new customers based on real‑time interest and movement.

Smart Payment: Enable facial‑recognition‑based checkout for faster, contactless transactions.

Smart Forecasting: Adjust product placement and inventory based on daily, weekly, and monthly customer interest patterns.

Smart Security: Monitor high‑risk individuals, stairways, lighting, and fire hazards to ensure safety.

By leveraging Suning’s big‑data capabilities and visual algorithms, Beidou injects a “smart brain” into retail stores, delivering data‑driven, personalized experiences and transforming offline retail into an intelligent, responsive ecosystem.