在人工智慧日益普及下，透過深度學習技術能根據影像特徵對物件進行辨識、分類和偵測，藉此作為異常事件檢測的依據，然而在大規模監視影像分析中，異常事件的目標對象只佔一小部分，因此將所有監視影像上傳到伺服器進行深度學習的物件偵測時會浪費大量上行頻寬來進行傳輸而導致頻寬不足的情形，間接限制了上傳至伺服器影像的數量，因此為了解決上述問題，本研究提出了以多層過濾影像機制降低上行頻寬使用量之研究，藉由第一層的前景物件過濾器快速過濾出包含前景物件的影像來降低目標過濾器需要偵測的影像數量，第二層目標過濾器再從過濾出的影像中，透過物件偵測模型過濾出包含目標物件的影像再上傳至伺服器，藉由丟棄不包含目標物件的影像來降低上傳資料量。此外，為了提升多個過濾器的執行效率，採用多執行緒來提升過濾器執行效率以及使用 Blocking Queue 來同步過濾器之間處理速度，實驗結果發現在前景物件過濾器中使用動態過濾閾值能降低過濾錯誤率的產生，加上目標物件過濾器後能再降低前景物件過濾器造成的過濾錯誤，最後採用本研究所提出的機制，能降低約79％上行頻寬使用量同時維持2％以下過濾錯誤率。

With the increasing popularity of artificial intelligence, deep learning techniques can be used to identify, classify, and detect objects based on image features, which can be used as the basis for abnormal event detection. However, according to surveillance image analysis, the target objects in abnormal events only account for a small part. Uploading all the surveillance images to server for object detection waste a large amount of uplink bandwidth, which indirectly reduces the number of images uploaded to the server for detection. In order to solve the problems above, the research proposes a multi-layer filtering image mechanism to reduce the usage of uplink bandwidth. The first layer of foreground object filter quickly filters out images containing foreground objects to reduce the number of images that target object filter needs to identify. The second target filter filters out the images that do not contain the target object through the object detection model to reduce the amount of uploaded data. In addition, the research adopt multi-thread implementation to improve the execution efficiency of multiple filters and Blocking Queue to synchronize the processing speed between filters. The results found that using dynamic filter threshold in the foreground object filter reduces the filtering error rate. In addition, adding additional target object filter can reduce the filtering error caused by the foreground object filter. In consequence, the mechanism proposed by this research can reduce the usage of uplink bandwidth by about 79％ while maintaining the filtering error rate below 2％.

[1] Q. Wang, J. Wan, and Y. Yuan, “Locality constraint distance metric learning for traffic congestion detection,” Pattern Recognition, Distance Metric Learning for Pattern Recognition, vol. 75, pp. 272–281, Mar. 1, 2018, issn: 0031-3203. doi: 10. 1016/j.patcog.2017.03.030. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S0031320317301401 (visited on 06/15/2021).
[2] chtseng. “OpenCV 的前景/背景分離技術,” CH.Tseng. (Nov. 3, 2018), [Online]. Available: https://chtseng.wordpress.com/2018/11/03/opencv%e7%9a%84% e5%89%8d%e6%99%af%ef%bc%8f%e8%83%8c%e6%99%af%e5%88%86%e9%9b%a2%e6% 8a%80%e8%a1%93/ (visited on 06/13/2021).
[3] P. KaewTraKulPong and R. Bowden, “An improved adaptive background mixture model for real-time tracking with shadow detection,” in Video-Based Surveillance Systems, P. Remagnino, G. A. Jones, N. Paragios, and C. S. Regazzoni, Eds., Boston, MA: Springer US, 2002, pp. 135–144, isbn: 978-1-4613-5301-0 978-1-4615- 0913-4. doi: 10.1007/978-1-4615-0913-4_11. [Online]. Available: http://link. springer.com/10.1007/978-1-4615-0913-4_11 (visited on 06/13/2021).
[4] Z. Zivkovic, “Improved adaptive gaussian mixture model for background subtrac- tion,” in Proceedings of the 17th International Conference on Pattern Recogni- tion, 2004. ICPR 2004., ISSN: 1051-4651, vol. 2, Aug. 2004, 28–31 Vol.2. doi: 10.1109/ICPR.2004.1333992.
[5] Z. Zivkovic and F. van der Heijden, “Efficient adaptive density estimation per image pixel for the task of background subtraction,” Pattern Recognition Letters, vol. 27, no. 7, pp. 773–780, May 1, 2006, issn: 0167-8655. doi: 10.1016/j.patrec. 2005.11.005. [Online]. Available: https://www.sciencedirect.com/science/ article/pii/S0167865505003521 (visited on 06/13/2021).
[6] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 779–788. [Online]. Available: https://www.cv-foundation.org/openaccess/content_cvpr_2016/html/ Redmon_You_Only_Look_CVPR_2016_paper.html (visited on 06/13/2021).
[7] J. Redmon and A. Farhadi, “YOLO9000: Better, faster, stronger,” arXiv:1612.08242 [cs], Dec. 25, 2016. arXiv: 1612.08242. [Online]. Available: http://arxiv.org/ abs/1612.08242 (visited on 06/13/2021).
[8] ——, “YOLOv3: An incremental improvement,” arXiv:1804.02767 [cs], Apr. 8, 2018. arXiv: 1804.02767. [Online]. Available: http://arxiv.org/abs/1804.02767 (visited on 06/13/2021).
[9] “GStreamer 系列 - 基本介绍 - John.Leng - 博客园.” (), [Online]. Available: https: //www.cnblogs.com/xleng/p/10948838.html (visited on 06/13/2021).
[10] See Jackson Hole, Town square webcam - northeast - SeeJH.com, Sep. 24, 2019. [Online]. Available: https://www.youtube.com/watch?v=ELtPf8zR8NY (visited on 06/18/2021).