臺灣博碩士論文加值系統

摘要

室內定位可以廣泛應用於各種應用場景,例如在大型室內場域中(例如車站、醫院、shopping mall、展覽場、博物館等)的行人交通指引，人員定位服務，人流追蹤等等。此外在定位功能上也可分為主動式(由系統主動偵測，例如人流追蹤)與被動式(由使用者主動提出，例如使用者定位服務)。目前研究人員研發不同可能的室內定位技術，大多數的室內定位技術使用無線電波，且多半都需要傳感器（sensor）支援且皆要在意收訊的角度、收訊時間差、訊號強度等成本問題。本論文針對大型車站(例如台北車站)室內定位問題提出一個以車站現有設施如行人路標指示牌認識為基礎的室內定位系統，使用者僅須使用其隨身攜帶的智慧型手機(具備上網功能)即可使用。本系統具有高可用性(high availability) 、方便性 (high accessibility and easy to use) 、低建置成本(low deployment cost) 、 高定位訊號穩定習性(robust sensor signal quality)等優點。本系統的核心模組是建立在一個具有高辨識準確度的行人路標指示牌認識系統。我們利用Google的Tensorflow架構平台，使用faster rcnn resnet101深度學習訓練模型做為行人路標指示牌辨識系統。在綜合三類由不同人、在不同時間、使用不同手機、以不同位置與角度拍攝，共6,341張測試影像測試下，綜合不同的訓練模型參數所得的不同訓練模型測試結果，平均的正確率可達98.3%(0.982930)。其中在更接近真實使用者的測試模擬結果其平均準確度更高達0.988746，這個結果讓我們相信我們所提的方法與系統架構是實際可行的。
Google Tensorflow架構平台是目前非常受歡迎的人工智慧開發平台，雖然我們使用的是現有的Google Tensorflow架構平台，但如何在這開發平台上提高行人路標指示牌的辨識準確度，有賴進一步對系統運用程序與相關系統參數做詳細分析與探討，本論文另一個重點即是對這些系統設計問題這深入探討。此外本論文雖以台北車站為案例探討，我們認為同樣的技術可以套用到其他類似的大型車站。 同時我們也認為同樣的技術可以運用到其他室內定位情境中，例如醫院、博物館、展場、遊樂場等等。因為這些大型的公共服務場所，為有效的導引旅客人潮到達他們要去的地點或措施，通常都同樣會設置明顯的導引路標指示牌。另外，我們也相信同樣的系統架構下，相同的技術也可運作的其他不限於路標指示牌的物件，例如對於博物館中的展覽品，特殊室內設施等等。

關鍵詞：機械學習、影像辨識、室內定位。

Abstract

Indoor positioning technologies can be applied to many use cases in our daily life, for example visitor guidance for directions and information in a public service facility such as train stations, hospitals, shopping malls, and museums, to name a few. Due to the great demands of indoor positioning services, researches have been trying to proposal different solutions to effectively and efficiently provide the services in different aspects. Among many others, indoor positioning services based on radio signals, such as WIFI, Bluetooth, and RFID, have attracted most interest in previous works. However, due to characteristics of radio signal propagation, the radio signal used for the positioning information could be interfered significantly by a large number of passengers/visitors in a hot spot, such as a busy train station as the case we would like to study in this thesis.
In this thesis, we propose a new design strategy to provide indoor positioning service based on the situation in a large train station such as Taipei Train Station, in which we propose to use signage recognition as the basis for the indoor positioning service. Compared with those approaches based on radio signals, the benefits are manifold with the proposed scheme. It provides high availability, high accessibility, and easy to use. Moreover, the deployment cost is nearly free, and it is more robust to a large number of visitors/passengers such as for the case as in a busy train station.
The core modules of the proposed indoor positioning system rely on a high precision signage recognition system. We use Google Tensorflow architecture platform to develop the signage recognition subsystem. How to develop a high precision signage recognition system has been a core design issue discussed in this thesis study. The effectiveness of the proposed signage recognition system is justified by large number of test images. There are totally 52 different signage are considered in the field trial (available in B1 core area of Taipei Train Station). Three test image sets, adding up to 6,341 images, are collected at different time, from different users, and with different models of smartphones. Our experiments show that in average about 98.3% (0.982930) of the 6,341 test images are correctly classified from 52 candidate signage. The high accuracy signage recognition rate justifies our proposed scheme to use recognition of pedestrian directional signage as the mean for indoor positioning services.


Keywords: mechanical learning, image recognition, indoor positioning.

目錄
第一章 緒論 1
1.1研究背景、動機與主題 1
1.2論文主要貢獻 3
1.3論文架構 5
第二章 文獻探討及相關技術 6
2.1 Tensorflow 深度學習開發框架 6
2.2 深度學習網路模型架構 9
第三章 系統設計與實作 12
3.1 Tensorflow Object Detection API 安裝 12
3.2 辨識模型(detection model)建立 15
3.2.1 Training image collection scheme 15
3.2.2 訓練模型 15
3.2.3 準備Pascal VOC 資料檔 18
3.2.4 Pascal_label_map 資料描述檔 23
3.2.5 產生TFRecord格式訓練資料檔 24
3.2.6 其餘training所需要的資料 25
3.3 訓練模型的資料檔 27
3.3.1 Training 運行指令 28
3.3.2 Tensorboard訓練可視化工具 29
3.3.3 模型輸出 30
第四章 系統測試與結果 32
4.1系統測試相關準備及處理作業 32
4.1.1 Testing data準備 32
4.1.2測試照片處理 33
4.1.3測試使用的程式 35
4.1.4本實驗之測試規劃表 35
4.2 Flip-On模型測試結果 36
4.2.1 dataSet1測試結果 36
4.2.2 dataSet2測試結果 37
4.2.3 dataSet3測試結果 39
4.2.4 Flip-ON測試結果討論 40
4.3 Flip-off模型測試結果 42
第五章 實驗結果分析 46
5.1 不同訓練模型參數選擇對正確辨識率的影響 46
5.1.1 使用不同數量的training data的差異 46
5.1.2 使用不同手機的測試照片的差異 47
5.1.3 模型使用不同框法的差異 48
5.1.4 Flip-ON與 Flip-Off 選項的差異 49
5.2 dataSet2辨識錯誤可能因素探討 50
5.2.1 dataSet2/Flip-off主要辨識錯誤原因分析 50
5.2.2 dataSet2/Flip-off主要辨識錯誤原因分布 54
5.3 dataSet1 與 dataSet3辨識錯誤可能因素探討 58
5.3.1 dataSet1/Flip-off 主要辨識錯誤原因分析 58
5.3.2 dataSet3/Flip-off 主要辨識錯誤原因分析 60
第六章 結論與未來展望 65
6.1研究結果摘要與結論 65
6.2未來展望 68
參考文獻 69
附錄 71
附錄A: Flip-ON 測試結果 75
附錄B: Flip-OFF測試結果 95
附錄C: dataSet2 test image 誤判圖檔 106
附錄D: 模型訓練時的validation loss資訊 121

參考文獻

1. TensorFlow dataflow graph. Online available: https://www.tensorflow.org/guide/graphs
2. TensorFlow Object Detection API - what do the losses mean in the object detection api? Online available : https://stackoverflow.com/questions/48111847/tensorflow-object-detection-api-what-do-the-losses-mean-in-the-object-detectio
3. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun,"Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks", in Proceedings of Neural Information Processing Systems 2015 (NIPS 2015). online avaialbe: https://arxiv.org/abs/1506.01497 (2015)
4. Karen Simonyan and Andrew Zisserman, "Very Deep Convolutional Networks for Large-Scale Image Recognition", online available: https://arxiv.org/abs/1409.1556
5. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun,"Deep Residual Learning for Image Recognition", IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770-778.
6. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, "Identity Mappings in Deep Residual Networks", online available: https://arxiv.org/abs/1603.05027
7. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, "Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition", online available: https://arxiv.org/abs/1406.4729
8. Ross Girshick, "Fast R-CNN", in Proceedings of 2015 IEEE International Conference on Computer Vision (ICCV), 2015.
9. Ross Girshick, Jeff Donahue1, Trevor Darrell1, and Jitendra Malik, "Rich feature hierarchies for accurate object detection and semantic segmentation", in Proceedings of the IEEE conference on computer vision and pattern recognition (ICCV), 2014.
10. Joyce Xu, "Deep Learning for Object Detection: A Comprehensive Review", Towards Data Science (posted on Sep. 11, 2017). Online available: https://towardsdatascience.com/deep-learning-for-object-detection-a-comprehensive-review-73930816d8d9
11. 深度學習目標檢測模型全面綜述：Faster R-CNN、R-FCN和SSD. https://www.itread01.com/content/1546131794.html
12. G. T. Wang, "TensorFlow Object Detection API 自行訓練模型教學，辨識特定物件", Online available: https://blog.gtwang.org/programming/ tensorflow-object-detection-api-custom-object-model-training-tutorial/2/
13. G. T. Wang, "TensorFlow Object Detection API 自動辨識物件教學", Online available: https://blog.gtwang.org/programming/tensorflow-object-detection-api-tutorial/
14. 木刻思專欄, "RUN FASTER RCNN USING TENSORFLOW DETECTION API", Online available: https://data-sci.info/2017/06/27/run-faster-rcnn-tensorflow- detection-api/
15. Priya Dwivedi, "Building a Toy Detector with Tensorflow Object Detection API", online available: https://towardsdatascience.com/ building-a-toy-detector-with-tensorflow-object-detection-api-63c0fdf2ac95
16. Dat Tran, "How to train your own Object Detector with TensorFlow’s Object Detector API", online available: https:// towardsdatascience.com/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9
17. Jiancheng Li, "使用TensorFlow Object Detection API进行物体检测", online available: https://lijiancheng0614.github.io/2017/08/ 22/2017_08_22_TensorFlow-Object-Detection-API/
18. The PASCAL (Pattern Analysis, Statistical Modelling and Computational Learning) VOC (Visual Object Classes) project. online available: http://host.robots.ox.ac.uk/pascal/VOC/
19. The PASCAL Visual Object Classes Challenge 2007. online available: http://host.robots.ox.ac.uk/pascal/VOC/voc2007/index.htmlShaoqing
20. Colabeler: Best AI Annotation Tool Ever. online available: http://www.colabeler.com/
21. 殘差網路(Residual Network). online available: https://www.itread01.com/content/ 1544962578.html