基于视频的智能零售柜关键算法研究

发布时间：2021-01-24 13:29

　　智能零售店已经成为吸引许多科技公司的关注的活跃话题。亚马逊、深兰科技奥兰治和IBM等公司已经开始通过搭建完整的智能零售店或零售柜来提高客户的购物体验。这些公司正在使用较为复杂的集成系统来实现其目标。但是,在考虑具体搭建一个小规模的智能零售店场景或零售柜时,首先需要解决一些关键问题:如何生成和标记属于某个域的所需图像数据集,比如,拟售商品在不同视角下的拍照图像;设计用于所售商品检测识别机器学习轻量化模型,它要同时满足精度、速度和存储容量等多个方面的要求;设计基于客户人脸图像的年龄估计和性别识别的轻量化机器学习模型。本文在较为全面地整理归纳智能零售店和零售柜的基础上,提出了一款智能零售柜总体设计方案,并就其中的三个关键问题,进行较为深入的研究,提出并实现了相应的解决方案。为了生成用于深度学习训练的带标注的图像数据集,本文设计了一条简单高效的人机协同的处理流水线。首先,对于每一类拟出售的商品,通过人工拍照或其它渠道,采集得到包含有该商品的图像,并赋予商品的类标属性,形成初步的图像数据集。随后,从每类商品中随机选择小部分图像,采用预训练的Mask RCNN模型,生成可疑目标的边界外框,经过人工... 

【文章来源】：厦门大学福建省 211工程院校 985工程院校 教育部直属院校

【文章页数】：96 页

【学位级别】：硕士

【文章目录】：
摘要
Abstract
Chapter 1 Introduction
    1.1 Future of Smart Retail Market
    1.2 Challenges
        1.2.1 Image Data Annotation
        1.2.2 Design Lightweight Model
        1.2.3 Base Features for Product Recommendation
    1.3 Related Work
        1.3.1 Recent Work in Smart Retail Stores
        1.3.2 Literature for Product Detection and Product Classification
        1.3.3 Related Work in Age and Gender Estimation for ProductRecommendation
    1.4 Our Solutions for the Challenges
        1.4.1 Training Image Dataset
        1.4.2 Naive Bounding Box Annotation Pipeline
        1.4.3 Product Detection
        1.4.4 Product Recommendation
    1.5 Contributions
    1.6 Structure of the Thesis
Chapter 2 System Architecture of the Smart Retail Cabinet
    2.1 Environment Designs for Smart Retail Cabinet
    2.2 System Diagram
    2.3 Logical Diagram
        2.3.1 Product Detection
        2.3.2 Product Recommendation
    2.4 Summary
Chapter 3 Naive Bounding Box Annotation
    3.1 Preparing Custom Image Dataset
    3.2 Bounding Box Annotation is a Challenge
    3.3 Overcome the Challenge
    3.4 Motivation
    3.5 Stage 1-Classification Models
        3.5.1 Preparing Cropped Images
        3.5.2 Features for Training Classification Model
        3.5.3 Image Augmentation
        3.5.4 Pre-Trained Feature Extraction Model
        3.5.5 Training the Classification Models
        3.5.6 Results for Classification Models
    3.6 Stage 2-Naive bounding Box Annotation Pipeline
        3.6.1 Pre-trained Object Detector
        3.6.2 Bounding Box Annotation Process
        3.6.3 Experiment for Naive Bounding Box Annotation Pipeline
        3.6.4 Results for Naive Bounding Box Annotation Pipeline
    3.7 Summary
Chapter 4 Product Detection in Smart Retail Cabinet
    4.1 One-Staged object detectors
    4.2 Product Detection Flow in Smart Retail Cabinet
    4.3 Evaluation Metric
    4.4 Object Detection Models for Smart Retail Cabinet
        4.4.1 SSD Models: MobileNetV2
        4.4.2 SSD Models: VGG16
        4.4.3 Tiny-YOLO V2
        4.4.4 Training and Experiment
        4.4.5 Results for Object Detection Models
    4.5 Lightweight Models for Smart Retail Cabinet
        4.5.1 MobileNet Architecture for YOLO algorithm
        4.5.2 Designing a Custom Module
        4.5.3 Custom Module inside YOLO model
        4.5.4 Custom Model Architecture
        4.5.5 Training Details for Custom Model
        4.5.6 Experimental Results
    4.6 Product Tracking in Smart Retail Cabinet
        4.6.1 Split Screen
        4.6.2 Movement Direction
        4.6.3 Counters
        4.6.4 Centroid Tracking Algorithm
        4.6.5 Drawbacks of the Tracking Algorithm
        4.6.6 Results with Product Tracking
    4.7 Summary
Chapter 5 Product Recommendation for Smart Retail Cabinet
    5.1 Adience Face Image Dataset
        5.1.1 Attributes of Adience Face Image Dataset
        5.1.2 Pre-processing
    5.2 Face Detection for Product Recommendation
    5.3 Face Detection Model in Smart Retail Cabinet
    5.4 Age and Gender Estimation in Smart Retail Cabinet
        5.4.1 Age and Gender estimation model
        5.4.2 Objective Function for Age and Gender Estimation
        5.4.3 Training the Models
        5.4.4 Results for Age and Gender Estimation
        5.4.5 Age Estimation-Precision Recall Curves
        5.4.6 Gender Estimation-ROC Curves
        5.4.7 Results on Test Images
        5.4.8 Results on Other Images
    5.5 Summary
Chapter 6 Conclusion and Future Work
    6.1 Summary
    6.2 Conclusions
    6.3 Future Improvements and Research Areas
Acknowledgement
References
Publications



本文编号：2997328