由於深度學習在電腦視覺上取得的成果，動作辨識的數據取得減少了對傳感器的依賴，從影片中取得資訊來分析人類的行為成為了另一個選項，並根據各種狀況的需求下，發展了不同的特徵資訊提取方式以及後續的數據處理。
隨著自駕車市場規模的成長，自動駕駛技術儼然成為目前研究的熱門方向。本論文的目的是將深度學習技術應用在自動駕駛車輛，當面對交通警察指揮時的情境下如何對交通指揮手勢進行辨識，並針對從不同方向的車道視角以及模擬可能出現的視角遮擋情況對辨識結果造成的影響進行探討。
所使用的方法基於人體骨架姿態，以關節點檢測網路對影片進行關節點提取，並進一步計算出關節之間的長度及夾角作為空間特徵以及透過時間序列網路提取時序特徵，最後獲得指揮手勢的辨識結果。在影片幀率為30幀的情況下，前方視角的準確率為91.56%，整體準確率為78.88%。

Due to the achievements made by deep learning in computer vision, the data acquisition of action recognition less relies on additional sensor devices. Retrieving information from videos to analyze human behavior has become another option. According to the needs of various situations, different feature information extraction methods and subsequent data processing methods have been developed.
With the growth of the self-driving market, self-driving technology has become a hot research direction. The purpose of this paper is to apply deep learning technology to recognize traffic command gestures in the context of autonomous vehicles, when facing traffic police commands. The influence of the viewpoints of the lanes from different directions and the possible blocking of the viewpoints on the identification results are discussed.
The method that is applied in this paper is based on the posture of the human skeleton, and then follows the steps: Extract the joint points from the video by a joint point detection network; calculate the length and angle between the joints as spatial features; extract the time series features through a time series network. Finally, the recognition results of command gestures are obtained. When the frame rate of the film is 30 frames, the accuracy rate of the front view angle is 91.56%, and the overall accuracy rate is 78.88%.

第1章 前言 1
1-1 研究背景與動機 1
1-2 研究目的 3
1-3 論文綱要 3
第2章 相關研究 5
2-1 交通指揮介紹 5
2-1-1 手勢分類 5
2-1-2 台灣的交通指揮手勢 6
2-2 動作辨識的發展 7
2-2-1 機器學習 7
2-2-2 深度學習 9
2-2-3 研究分支 11
2-3 關節點檢測 12
2-3-1 相關數據集 12
2-3-2 算法介紹 13
第3章 交通指揮手勢辨識 15
3-1 交通指揮手勢辨識流程 15
3-2 數據介紹 16
3-3 OpenPifPaf 19
3-4 格式處理 22
3-5 空間特徵提取 24
3-6 時序網路模型 27
第4章 實驗方法與結果 29
4-1 實驗環境 29
4-2 評估方法 30
4-3 實驗方法討論 30
4-3-1 時序網路有無的比較 31
4-3-2 不同方向視角的比較 32
4-3-3 採樣幀率比較 32
4-3-4 序列長度比較 33
4-3-5 輸入比較 34
4-3-6 遮蔽模擬 35
第5章 結論與未來工作 41
5-1 結論 41
5-2 未來工作 41
參考文獻 43

交通號誌（無日期）。維基百科。取自2022年8月23日 https://zh.wikipedia.org/wiki/%E4%BA%A4%E9%80%9A%E8%99%9F%E8%AA%8C
自動駕駛汽車（無日期）。維基百科。取自2022年8月23日 https://zh.wikipedia.org/zh-tw/%E8%87%AA%E5%8B%95%E9%A7%95%E9%A7%9B%E6%B1%BD%E8%BB%8A
何堅, 王偉東, & 張丞（2020）。空間上下文與時序特徵融合的交警指揮手勢識別技術。電子學報， 48（5），頁 966-974。 doi: 10.3969/j.issn.0372-2112.2020.05.018
隋昱嬋（2021年11月10日）。手機鏡頭就能「骨骼標定」，靠AI揪錯誤姿勢！健身App Uniigym還有哪些亮點？。取自 https://www.bnext.com.tw/article/66084/workout-ai-app-deeplearning
廖家宜（2019年3月14日）。人體行為辨識技術助勞力密集製造業提升生產效率。取自 https://www.digitimes.com.tw/iot/article.asp?id=0000555447_EQ93Z6VL5VXMQO662V06G
盧春男, & 陳存雄（2019年）。中華民國汽車駕駛人教育學會題庫。取自 https://wang.i234.me/123128/AS013/0004-JPG/index.htm
鑫豪科技（無日期）。跌倒偵測系統 – 鑫豪科技。取自 http://www.camels.com.tw/%E8%B7%8C%E5%80%92%E5%81%B5%E6%B8%AC%E7%B3%BB%E7%B5%B1/
Cubo Ai（無日期）。Cubo Ai Plus 智慧寶寶攝影機 - 第一台兼具安全、睡眠分析與時光牆的嬰兒監視器。取自 https://tw.getcubo.com/
openpifpaf. (n.d.). openpifpaf. Retrieved August 25, 2022, from https://github.com/openpifpaf/openpifpaf
SAE INTERNATIONAL. (n.d.). Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles. Retrieved August 23, 2022, from https://www.sae.org/standards/content/j3016_202104/
Elias, P., Sedmidubsky, J., & Zezula, P. (2019). Understanding the Gap between 2D and 3D Skeleton-Based Action Recognition. Proceedings of the IEEE International Symposium on Multimedia, pp. 192-1923. doi.10.1109/ISM46123.2019.00041
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770-778. doi.10.1109/CVPR.2016.90
Kreiss, S., Bertoni, L., & Alahi, A. (2019). PifPaf: Composite Fields for Human Pose Estimation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11969-11978. doi.10.1109/CVPR.2019.01225
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet classification with deep convolutional neural networks. Commun. ACM, 60(6), 84–90. doi: 10.1145/3065386
Papandreou, G., Zhu, T., Kanazawa, N., Toshev, A., Tompson, J., Bregler, C., & Murphy, K. (2017). Towards Accurate Multi-person Pose Estimation in the Wild. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3711-3719. doi.10.1109/CVPR.2017.395
Tran, D., Bourdev, L., Fergus, R., Torresani, L., & Paluri, M. (2015). Learning Spatiotemporal Features with 3D Convolutional Networks. Proceedings of the IEEE International Conference on Computer Vision, pp. 4489-4497. doi.10.1109/ICCV.2015.510
Wang, H., & Wang, L. (2018). Beyond Joints: Learning Representations From Primitive Geometries for Skeleton-Based Action Recognition and Detection. IEEE Transactions on Image Processing, 27(9), 4382-4394. doi: 10.1109/TIP.2018.2837386
Wei, S. E., Ramakrishna, V., Kanade, T., & Sheikh, Y. (2016). Convolutional Pose Machines. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4724-4732. doi.10.1109/CVPR.2016.511
Wiederer, J., Bouazizi, A., Kressel, U., & Belagiannis, V. (2020). Traffic Control Gesture Recognition for Autonomous Vehicles. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 10676-10683. doi.10.1109/IROS45743.2020.9341214
Wu, Y., Qiu, H., Wen, J., & Feng, R. (2020). OSD: An Occlusion Skeleton Dataset for Action Recognition. Proceedings of the IEEE International Conference on Big Data, pp. 3355-3360. doi.10.1109/BigData50022.2020.9378458