对于现有的无人零售技术的调研

1. 深度解读Amazon go核心技术，人加智能助力无人零售

   • 国外：亚马逊

   • 国内：阿里、京东

   • “just walk out”概念，这个概念对超市场景来说主要是减少了两个环节：①货品扫码；②货品付款，也就是结账环节。 所谓的“无人便利店”。其实是有服务人员的，有人上货，有人做即时性的食品（奶茶等等），准确来说amazon go做的是的自动结算系统（cashierless），没有收银员， 「拿了就走」 ，系统自动识别顾客拿走的物品，自动结账。

   • 解决的问题： 结账环节虽然非常关键，但对店家来说是一个鸡肋环节，因为大部分的工作都是重复化、机械化的操作，如收银的扫码，结算等。这类工作不需要有很高的技术水平就可以胜任，薪资水平一直也比较低。却需要大量员工操作，人力成本、设备成本较高； amazon go正是解决了这个环节。

   • 真实意图：

     • 数据收集技术，可以将消费者入店的所有行为数据进行跟踪与分析
     • 希望能更切中零售的本质：出售受消费者喜爱的产品。

   • 我国自动结算零售的发展历程：

     • 自动结算1.0时代：自助结算机

       将收银的工作转移给用户，结账的时候必须要有专人检查，否则货损问题无法解决。但专人检查效率低，提防性行为带给消费者的感受不好。

     • 自动结算2.0时代： RFID

       产品外贴设置好的RFID标签，顾客挑选完毕后，会到读取设备上读取自己所购买的商品，并进行支付。成本高，容易损坏。

     • 自动结算2.5时代：2D视觉识别+RFID

     • 无人值守3.0

       RGBD成像技术：RGBD相机取代传统的2D摄像头

       边缘计算：配备了边缘计算方案，开发边缘计算平台。RGBD相机通过多种接口进行深度对接，如网络、USB3.0和MIPI接口。

       人体跟踪技术：打造了多相机头顶阵列传感方案，解决了多相机的时间同步、空间标定、安装施工难度大等一系列难题，从传感方案角度最大可能的保证了人体跟踪效果。 采用RGB-D跟踪技术，为物体建立三维坐标，从而彻底的解决了跟踪困难与后期识别的问题。

2. How the Amazon Go Store’s AI Works

   6 core problems of "Computer Vision Complete" problem:

   • Sensor Fusion
   • Calibration 校准
   • Person detection
   • Object Recognition
   • Pose estimation
   • Activity Analysis
     

   Person Identification

   • Locator

     Problems: Occlusion(遮挡，人被物体遮挡), Tangled State(纠缠态，人和人太密集)

     Amazon Solution: uses custom camera hardware that does both RGB video and distance calculation. They segment image into pixels, group pixels into blobs, and label each blob as person/not-person. Finally, they build a location map from the frame using triangulation of each person across multiple cameras.

   • Linker

     To tracking the customers in the store

   • Disambiguating Tangled States 消除纠缠态

     Mark close customers as low confidence get scheduled to be re-identified over time.

     Distinguish associates(员工) from customers.

   Item identification

   • Product ID detection

   • Customer association

     Combining all of the information from the above to finally answer the “Who took what?” question.

   • Pose Estimation

     Build a stick-figure like model of the customer from the video, because cameras look from the top down, not form an isometric view.

   • Action determination

     The system needs to count all the items on the shelf rather than using a simple assumption based on space.
     

   • The long tail

     Using simulation to build a massive training set

   Streaming Services

   1. Video capture with compute on board to do basic preprocessing and cut down the bandwidth requirements
   2. Video streamer appliance on site to handle video codecs, network issues, and guarantee delivery to the cloud
   3. Video servers on the cloud to capture and store video in S3 and Dynamo

   Entry & Exit Detection

   1. Mobile App to scan QR when you show up at the store. They spent a lot of time doing UX testing on this (scan with phone up or down, how to handle groups, etc.)
   2. Association System associates your likeness in the video to your account based on position in the store entrance when you scan the QR code
   3. Creation of the session happens based on the association

   问题:1. 重复扫码 2. 家庭购物(group 一人支付)

   Cart, Payment, and Receipts