跳到主要內容

臺灣博碩士論文加值系統

(44.210.149.205) 您好!臺灣時間:2024/04/12 21:45
字體大小: 字級放大   字級縮小   預設字形  
回查詢結果 :::

詳目顯示

我願授權國圖
: 
twitterline
研究生:賴頎鈞
研究生(外文):Chi-Chun Lai
論文名稱:可應用於智慧型自走載具且基於YOLO深度學習技術的顧客導引系統
論文名稱(外文):Customer Guidance System with YOLO-based Deep Learning Technology for Smart Autonomous Mover
指導教授:范志鵬范志鵬引用關係
指導教授(外文):Chih-Peng Fan
口試委員:陳冠宏黃穎聰
口試委員(外文):Kuan-Hung ChenYin-Tsung Hwang,
口試日期:2022-07-21
學位類別:碩士
校院名稱:國立中興大學
系所名稱:電機工程學系所
學門:工程學門
學類:電資工程學類
論文種類:學術論文
論文出版年:2022
畢業學年度:110
語文別:中文
論文頁數:71
中文關鍵詞:嵌入式平台深度學習機器學習人臉識別色彩空間
外文關鍵詞:Embedded platformdeep learningmachine learningface recognitioncolor space
相關次數:
  • 被引用被引用:0
  • 點閱點閱:166
  • 評分評分:
  • 下載下載:0
  • 收藏至我的研究室書目清單書目收藏:0
近年來,隨著人工智慧應用的普及和硬體方面的成熟,越來越多設計將深度學習與嵌入式平台結合,智慧型自走載具就是其中一項熱門應用,如何在人群中有效識別顧客變成一個非常重要的研究課題。

本論文透過嵌入式平台(Jetson Nano),使用了基於深度學習和機器學習的人臉識別來判別顧客,並使用演算法來提取衣服的bounding box,將衣服的bounding box利用色彩空間來提取特徵並與顧客的資料互相比對,在抓取人臉bounding box時,同時使用性別分類器來進行性別識別,來判斷目前追蹤目標是否為顧客的註冊性別。最後,當三種條件經過系統判斷後,系統將會決斷當前是否為顧客。

為了了解各個方法的測試成效,在人臉檢測部分調用YOLOv3、YOLOv4、YOLOv3-tiny和YOLOv4-tiny並修改參數與輸入影像大小以便找到最佳解。在人臉識別方面採用了KNN和SVM機器學習,目的希望能找到能夠有不錯的準確率和處理速度。在衣服識別中採用了三種色彩空間,分別為HSV、HLS和CIELA,來萃取衣服特徵並比較三種方法中適合在最多種類別中不受光影變化影響最後,將三種條件組合成系統識別,並調整權重以達到最佳解。

實驗結果顯示,兩票式版本的FPS和重新訓練時間分別為6.0和235秒,以及其Accuracy、Precision、Recall和F1-Score分別為0.971、1.000、0.942和0.970。而透過減少5000張路人資料集,以及改善衣服識別和新增性別辨識的方式。在三票式的FPS和重新訓練時間,FPS降低 1.0,其原因是因為增加性別識別,但重新訓練時間大幅減少129秒。另外,將路人資料及刪減到50張並使用SVM來處理的三票式優化版本,其FPS跟兩票式版本提升0.7FPS以及重新訓練時間減少205秒,此方案為三種版本中表現最好。
In recent years, with the popularity of artificial intelligence applications and hardware development, more and more designs are combining deep learning and embedded platforms, and the intelligent autonomous vehicle is one of the popular applications, how to effectively identify customers in the crowd has become a very important research topic.

The embedded platform (Jetson Nano) is based on deep learning and machine learning face recognition to identify customers, and uses an algorithm to extract the bounding box of clothes, using the color space to extract the features of the bounding box of clothes and compare it with the customer's data. The gender classifier is used for gender recognition to determine whether the current tracking target is the registered gender of the customer. Finally, when the three conditions are judged by the system, the system will decide whether the current customer is a customer.

In order to understand the effectiveness of each method, in the face detection part, YOLOv3, YOLOv4, YOLOv3-tiny and YOLOv4-tiny were used and the parameters and input image size were modified in order to find the best solution. In face recognition, KNN and SVM machine learning are used in order to find a good accuracy and processing speed. Three color spaces, HSV, HLS, and CIELA, are used in clothing recognition to extract clothing features and compare the three methods that are suitable in the most categories without the effect of light and shadow changes . Finally, the three conditions are combined into a systematic identification and the weights are adjusted to achieve the best solution.

The experimental results show that the two-vote version has an FPS of 6.0 and a read time of 235 seconds, as well as Accuracy, Precision, Recall were respectively 0.971, 1.000, 0.942, and 0.970. And by reducing the number of passerby datasets by 5,000, as well as improving clothing recognition and adding gender recognition. In terms of FPS and read time for the three-vote, In terms of FPS and read time for the three-vote format, the FPS was reduced by 1.0 due to the addition of gender recognition, but the read time was significantly reduced by 129 seconds. In addition, the optimized three-vote version, which reduced the number of people data to 50 and used SVM, improved the FPS by 0.7 FPS and reduced the read time by 205 seconds compared to the two-vote version, which was the best performance among the three versions.
論文摘要 i
Abstract ii
目錄 iv
表目錄 vi
圖目錄 viii
1.緒論 1
1.1 研究動機與目的 1
1.2 系統架構 2
1.3 論文架構 3
2.文獻探討 4
2.1 基於Adaboost的人臉偵測 4
2.2 基於HOG的人臉偵測 5
2.3 基於MTCNN的人臉偵測 6
2.4 基於HOG使用dlib數據庫的人臉識別 7
2.5 基於MTCNN使用Facenet的人臉識別 8
3.預備知識 9
3.1 反向傳播 ( Backpropagation ) 9
3.2 激活函數 ( Activation Function ) 10
3.2.1 Sigmoid 10
3.2.2 ReLU與Leaky ReLU 11
3.3 梯度下降法 (Gradient Descent,GD) 11
3.4 YOLO模型 14
3.5 批量標準化 (Batch Normalization,BN) 16
3.6 先驗框 (anchor box) 16
3.7 損失函數 (Loss function) 18
3.8 多尺度訓練 (Multi-Scale Training) 18
3.9 YOLOv3與YOLOv3-tiny 19
3.10 YOLOv4與YOLOv4-tiny 20
4.演算法流程 22
4.1 資料集(dataset) 22
4.1.1 訓練資料集 22
4.1.2 LFW人臉數據庫 23
4.2 系統流程 23
4.3 人臉偵測 25
4.3.1 人臉偵測的標註檔(label) 25
4.3.2 LabelImg 26
4.3.3 資料集分割 28
4.3.4 建構YOLO訓練環境與參數調整 28
4.3.5 推論加速 30
4.4 人臉識別 31
4.5 衣服偵測 33
4.6 衣服識別 34
4.7 性別識別 39
4.8 顧客識別 39
5.實驗結果與分析 41
5.1 硬體訓練設備與測試環境 41
5.2 測試資料集 42
5.3 模型評估指標 42
5.3.1 Accuracy、Precision、Recall與F1-Score 43
5.3.2 TPR、FNR、FPR與TNR 44
5.3.3 AP與mAP 44
5.4 YOLO模型選擇與推論加速 45
5.5 兩票式顧客識別系統測試結果 47
5.5.1 人臉識別結果 47
5.5.2 衣服識別結果 50
5.5.3 顧客識別結果 51
5.6 三票式顧客識別系統測試結果 53
5.6.1 人臉識別結果 53
5.6.2 衣服識別結果 55
5.6.3 性別識別結果 56
5.6.4 顧客識別結果 57
5.7 三票式優化測試結果 60
5.7.1 人臉識別結果 60
5.7.2 衣服識別結果 61
5.7.3 顧客識別結果 62
6.結論與未來展望 67
6.1 結論 67
6.2 未來展望 68
參考文獻 69
[1] “人臉偵測 Face Detection 算法” [Online]. Available: https://reurl.cc/1Z5gAV
[2] “圖像特徵工程:HOG特徵描述子介紹” [Online].Available: https://reurl.cc/yEvm4q
[3] N. Zhang, J. Luo and W. Gao, "Research on Face Detection Technology Based on MTCNN," 2020 International Conference on Computer Network, Electronic and Automation (ICCNEA), 2020, pp. 154-158, doi: 10.1109/ICCNEA50255.2020.00040.
[4] H. Xia and C. Li, "Face Recognition and Application of Film and Television Actors Based on Dlib," 2019 12th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), 2019, pp. 1-6, doi: 10.1109/CISP-BMEI48845.2019.8965869.
[5] E. Jose, G. M., M. T. P. Haridas and M. H. Supriya, "Face Recognition based Surveillance System Using FaceNet and MTCNN on Jetson TX2," 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS), 2019, pp. 608-613, doi: 10.1109/ICACCS.2019.8728466.
[6] C. Szegedy et al., "Going deeper with convolutions," 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 1-9, doi: 10.1109/CVPR.2015.7298594.
[7] “機器學習-倒傳遞( Backward propagation)詳細推導” [Online]. Available: https://reurl.cc/plLDe
[8] “深度學習筆記——常用的激活(激勵)函數” [Online]. Available: https://codingnote.cc/zh-tw/p/176736/
[9] “梯度下降法” [Online].Available: https://blog.csdn.net/google19890102/article/details/50276775
[10] “Momentum示意圖” [Online].Available: https://zhuanlan.zhihu.com/p/34240246
[11] Redmon, J., Divvala, S., Girshick, R., "You Only Look Once-Unified Real-Time Object Detection," In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.779-788,2016.
[12] Redmon, J., & Farhadi, A. (2017). YOLO9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 7263-7271).
[13]S. Ren, K. He, R. Girshick, X. Zhang and J. Sun, "Object Detection Networks on Convolutional Feature Maps," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 7, pp. 1476-1481, 1 July 2017, doi: 10.1109/TPAMI.2016.2601099.
[14]“人臉偵測 Face Detection 算法” [Online]. Available: https://reurl.cc/4103jL
[15]“TensorRT 3: Faster TensorFlow Inference and Volta Support” [Online]. Available: https://developer.nvidia.com/blog/tensorrt-3-faster-tensorflow-inference/
[16] “A Quick Introduction to K-Nearest Neighbors Algorithm” [Online]. Available: https://medium.com/@adi.bronshtein/a-quick-introduction-to-k-nearest-neighbors-algorithm-62214cea29c7
[17] “支援向量機(Support Vector Machine)介紹” [Online]. Available: https://reurl.cc/NA7G4k
[18] “HSL和HSV色彩空間” [Online]. Available: https://reurl.cc/9GK1nx
[19] “Color Space vs. Color Tolerance” [Online]. Available: https://www.xrite.com/blog/tolerancing-part-3
[20] H. -L. Chen, C. -C. Lai, J. -M. Lin, K. -H. Chen, Y. -T. Hwang and C. -P. Fan, "Effective Two-Stage Processing Based Lite Deep Learning Classifier for Gender Detection," 2021 IEEE International Conference on Consumer Electronics-Taiwan (ICCE-TW), 2021, pp. 1-2, doi: 10.1109/ICCE-TW52618.2021.9602888.
[21] Chi-Chun Lai and Chih-Peng Fan, “Learning and Features based Customer Guidance Function for Application of Smart Autonomous Mover,” The 5th World Symposium on Communication Engineering (WSCE 2022), Nagoya University, Japan, September 2022. (EI)
[22] 賴頎鈞, 陳華倫, 林介民, 范志鵬, “基於YOLO深度學習技術的智慧型自走載具顧客導引模式系統設計與嵌入式軟體實現,” 2022 人工智慧技術及應用研討會 (Artificial Intelligence Technology and Application (AITA2022)) , 國立臺中科技大學, May 2022.
電子全文 電子全文(網際網路公開日期:20250801)
連結至畢業學校之論文網頁點我開啟連結
註: 此連結為研究生畢業學校所提供,不一定有電子全文可供下載,若連結有誤,請點選上方之〝勘誤回報〞功能,我們會盡快修正,謝謝!
QRCODE
 
 
 
 
 
                                                                                                                                                                                                                                                                                                                                                                                                               
第一頁 上一頁 下一頁 最後一頁 top
無相關期刊