由於棒球賽事的轉播在台灣受到廣泛喜愛，我們想要提供更多有關於當前比
賽的資訊來幫助觀眾對比賽有更深的了解或是幫助球隊的情蒐工作，特別是比賽
中投手們的投球球路。我們在這項研究中提出了一種識別賽事中投球類型的檢測
系統。所提出的方法將結合隱馬爾可夫模型和支持向量機對球路類別進行偵測。
在隱馬爾可夫模型的處理分支中，我們生成了呈現軌跡中方向變化的符號序列，
用以抓取不同球路類別的瞬時運動特性。隱馬爾可夫模型由各類別的訓練資料進
行學習，用於生成與各個隱馬爾可夫模型相對應類別的對數概率。在支持向量機
的處理分支中，首先檢測畫面中的參考點，並且對棒球軌跡進行標準化處理以減
少因為不同棒球賽事引起的拍攝角度或縮放的差異。由標準化後的軌跡生成許多
特徵進一步提取軌跡整體的運動特徵。我們進一步使用前饋式特徵選取的方法來
搜索降較好的特徵組合，之後使用支持向量機生成球路類別的對數概率。最後，
將這兩者的輸出進行軟投票的集成學習生成偵測結果。
為了驗證分類系統針對包含較多投手的資料集，仍保持較好的分類能力，我們在
2017 美國大聯盟季後賽中取出20 場的球賽生成了一個新資料集，包含了60 位
投手的投球軌跡共4,883 球。在我們的資料集中，針對以下五個投球類別進行標
記並分類：快速球，滑球，變速球，曲球和伸卡球。最後在新生成的資料集中，
我們提出的檢測系統可以達到的準確率為89.16%，五種類別的平均準確度可以
達到81.91%，其中三種球路的分辨準確率可以達到86.59%以上。

Because baseball broadcasts are so popular in Taiwan, we want to provide more information about the baseball games, especially the current pitching type of the ball thrown by the pitcher. A detection system that recognizes the pitching type in the baseball game broadcasting is proposed in this study. The proposed method will detect the pitching type by the combination of hidden Markov models (HMMs) and support vector machine (SVM).
In the processing branch of HMMs, we generate the symbol sequences representing the direction changes through the trajectories in order to capture the instantaneous temporal-dependent characteristics of different pitching types. HMMs are used to generate the log probability of the pitching type that corresponds to the class represented by each HMM. In the processing branch of SVM, the references in the frame are detected first and the normalization is applied to the trajectory to reduce the difference caused by different baseball games. Lots of features are generated from the normalized trajectory to extract overall motion characteristics. We further use the sequential forward selection to search out the better feature subset. Afterwards, the SVM is applied to generate the log probability of pre-defined pitching types. In the end, the detection result is generated by the soft voting from these two models.
To verify the classification ability on the data containing lots of pitchers, we generate a new dataset containing 4,883 pitching ball trajectories thrown by 60 pitchers in the 2017 MLB post-season. In our dataset, there are five pitching types labeled as below: fastball, slider, changeup, curveball, and sinker. With the proposed detection system on our new dataset, the accuracy can achieve 89.16%. The precision of five classes varies from 62.12% to 94.34% and the average precision is 81.91%.

Abstract . . . . . . . . . . . . . . . . . . . . . . . . i
1 Introduction . . . . . . . . . . . . . . . . . . . . . 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . 2
1.3 Main Contributions . . . . . . . . . . . . . . . . . 3
1.4 Organization . . . . . . . . . . . . . . . . . . . . 4
2 Related Works . . . . . . . . . . . . . . . . . . . . 5
2.1 Video Analysis in Sports . . . . . . . . . . . . . . 6
2.2 Baseball Video Analysis . . . . . . . . . . . . . . 8
2.2.1 Extraction of Baseball Trajectory . . . . . . . . 8
2.2.2 Event Detection . . . . . . . . . . . . . . . . . 12
2.2.3 Pitching Type Recognition . . . . . . . . . . . . 15
2.3 Comparison and Discussion . . . . . . . . . . . . . 17
3 Proposed Methodologies for Pitch Type Recognition . . 19
3.1 Trajectory Normalization . . . . . . . . . . . . . 21
3.1.1 References Detection . . . . . . . . . . . . . . 21
3.1.2 Normalization on the Trajectory . . . . . . . . . 26
3.2 Symbol Generation . . . . . . . . . . . . . . . . . 28
3.3 Hidden Markov Models . . . . . . . . . . . . . . . 30
3.3.1 Forward Algorithm . . . . . . . . . . . . . . . . 32
3.3.2 Forward-Backward Algorithm . . . . . . . . . . . 34
3.3.3 Pitch Type Recognition . . . . . . . . . . . . . 37
3.4 Feature Generation . . . . . . . . . . . . . . . . 38
3.4.1 Feature Extraction . . . . . . . . . . . . . . . 38
3.4.2 Feature Normalization . . . . . . . . . . . . . . 40
3.4.3 Feature Selection . . . . . . . . . . . . . . . . 42
3.5 Support Vector Machine (SVM) . . . . . . . . . . . 43
3.6 Soft Voting . . . . . . . . . . . . . . . . . . . . 47
4 Dataset Collection and Implementation Results . . . . 49
4.1 Dataset Collection . . . . . . . . . . . . . . . . 49
4.2 Feature Selection . . . . . . . . . . . . . . . . . 53
4.3 Experiment Result . . . . . . . . . . . . . . . . . 56
4.3.1 Classification Result for Our Dataset . . . . . . 56
4.3.2 Classification Result for Different Subsets . . . 57
4.3.3 Comparison with Literature . . . . . . . . . . . 62
5 Conclusion and Future Works . . . . . . . . . . . . . 65
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . 65
5.2 Future Works . . . . . . . . . . . . . . . . . . . 66
Bibliography . . . . . . . . . . . . . . . . . . . . . 67
Appendices . . . . . . . . . . . . . . . . . . . . . . 71
A Order List of Sequential Forward Selection . . . . . 73

[1] M. Fast, “What the heck is pitchf/x ?” Hardball Times Ann. 2010, pp. 153–158, 2010.
[2] K. Koseler and M. Stephan, “Machine learning applications in baseball: A systematic literature review,” Applied Artificial Intelligence, vol. 31, no. 9-10, pp. 745–763, 2017. [Online]. Available: https://doi.org/10.1080/08839514.2018.1442991
[3] A. Attarian, G. Danis, J. Gronsbell, G. Iervolino, and H. Tran, “A comparison of classification methods with an application to classifying baseball pitches,” 03 2013.
[4] P. Hoang, M. Hamilton, H. Tran, J. Murray, C. Stafford, L. Layne, and D. Padget, “Applying machine learning techniques to baseball pitch prediction,” 01 2014.
[5] J. Bock, “Pitch sequence complexity and long-term pitcher performance,” Sports, vol. 3, pp. 40–55, 03 2015.
[6] M. Takahashi, M. Fujii, and N. Yagi, “Automatic pitch type recognition from baseball broadcast videos,” 2008 Tenth IEEE International Symposium on Multimedia, pp. 15–22, 2008.
[7] C.-C. Li, C.-W. Lin, and J.-Y. Yu, “Statistical pitch type recognition in broadcast baseball videos,” Journal of Computers, vol. 21, no. 2, pp. 14–22, jul 2010.
[8] Y.-F. Huang and Z.-X. Yang, “Automatic player behavior analyses from baseball broadcast videos,” Active Media Technology, pp. 11–21, 2012.
[9] H. Shih, “A survey of content-aware video analysis for sports,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 28, no. 5, pp. 1212–1231, 2018.
[10] A. Gueziec, “Tracking pitches for broadcast television,” Computer, vol. 35, no. 3, pp. 38–43, 2002.
[11] G. Welch and G. Bishop, “An introduction to the kalman filter,” 1995.
[12] W. Chu, C. Wang, and J. Wu, “Extraction of baseball trajectory and physicsbased validation for single-view basebal video sequences,” 2006 IEEE International Conference on Multimedia and Expo, pp. 1813–1816, 2006.
[13] C. Huang, H.-S. Chen, M.-H. Hsiao, W.-J. Tsai, and S.-Y. Lee, “A trajectory-based ball tracking framework with visual enrichment for broadcast baseball videos.” J. Inf. Sci. Eng., vol. 24, pp. 143–157, 01 2008.
[14] B. Li and M. Ibrahim Sezan, “Event detection and summarization in sports video,” Proceedings IEEE Workshop on Content-Based Access of Image and Video Librarie(CBAIVL 2001), pp. 132–138, 2001.
[15] A. Watabe, K. Komiya, J. Usuki, K. Suzuki, and H. Ikeda, “Effective designation of specific shots on video service system utilizing mahalanobidistance,” IEEE Transactions on Consumer Electronics, vol. 51, no. 1, pp. 152–159, 2005.
[16] H. Chen, H. Chen, W. Tsai, S. Lee, and J. Yu, “Pitch-by-pitch extraction from single view baseball video sequences,” 2007 IEEE International Conference on Multimedia and Expo, pp. 1423–1426, 2007.
[17] Chih-Hao Liang, Jin-Hau Kuo, Wei-Ta Chu, and Ja-Ling Wu, “Semantic units detection and summarization of baseball videos,” The 2004 47th Midwest Symposium on Circuits and Systems, 2004. MWSCAS ’04., vol. 1, pp. I–297, 2004.
[18] M. Hung and C. Hsieh, “Event detection of broadcast baseball videos,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 18, no. 12, pp. 1713–1726, 2008.
[19] Peng Chang, Mei Han, and Yihong Gong, “Extract highlights from baseball game video with hidden markov models,” Proceedings. International Conference on Image Processing, vol. 1, pp. I–I, 2002.
[20] N. H. Bach, K. Shinoda, and S. Furui, “Robust highlight extraction using multistream hidden markov models for baseball video,” IEEE International Conference on Image Processing 2005, vol. 3, pp. III–173, 2005.
[21] H. Chen, C. Chou, W. Tsai, S. Lee, and B.-S. Lin, “Hmm-based ball hitting event exploration system for broadcast baseball video,” Journal of Visual Communication and Image Representation, vol. 23, no. 5, pp. 767–781, jul 2012.
[22] L. R. Rabiner, “A tutorial on hidden markov models and selecteapplications in speech recognition,” Proceedings of the IEEE, vol. 77, no. 2, pp. 257–286, 1989.
[23] C. Cortes and V. Vapnik, “Support-vector networks,” in Machine Learning, 1995, pp. 273–297.
[24] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector machines,” ACM Transactions on Intelligent Systems and Technology, vol. 2, pp. 27:1–27:27, 2011, software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.
[25] W.-N. Lie, G.-S. Lin, and S.-L. Cheng, “Pitching shot detection based on multiple feature analysis and fuzzy classification,” Advances in Multimedia Information Processing - PCM 2006, pp. 852–860, 2006.
[26] J. A. Hartigan and M. A. Wong, “Algorithm as 136: A k-means clustering algorithm,” Journal of the Royal Statistical Society. Series C (Applied Statistics), vol. 28, no. 1, pp. 100–108, 1979. [Online]. Available: http://www.jstor.org/stable/2346830
[27] R. Duda, P. Hart, and D. G.Stork, Pattern Classification. Wiley, 01 2001.
[28] E. I. Organick, A FORTRAN IV Primer. Addison-Wesley, 1966, p. 42, ”Some processors also offer the library function called ATAN2, a function of two arguments (opposite and adjacent)”.
[29] J. Bergstra and Y. Bengio, “Random search for hyper-parameter optimization,” Journal of Machine Learning Research, vol. 13, no. 10, pp. 281–305, 2012. [Online]. Available: http://jmlr.org/papers/v13/bergstra12a.html
[30] L. E. Baum, “An inequality and associated maximization technique istatistical estimation for probabilistic functions of Markov processes,” Inequalities III: Proceedings of the ThirSymposium on Inequalities, pp. 1–8, 1972.
[31] D. P. Muni, N. R. Pal, and J. Das, “Genetic programming for simultaneous feature selection and classifier design,” IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 36, no. 1, pp. 106–117, 2006.
[32] Chih-Wei Hsu and Chih-Jen Lin, “A comparison of methods for multiclass support vector machines,” IEEE Transactions on Neural Networks, vol. 13, no. 2, pp. 415–425, 2002.
[33] Cristianini, Nello, and S.-T. John, “An introduction to support vector machines and other kernel-based learning methods. repr,” Introduction to Support Vector Machines and other Kernel-Based Learning Methods, vol. 22, 01 2001.
[34] S. Tong and D. Koller, “Support vector machine active learning with applications to text classification,” The Journal of Machine Learning Research, vol. 2, pp. 45–66, 12 2001.
[35] S. Raschka, Python Machine Learning. Birmingham, UK: Packt Publishing, 2015.
[36] A. Piergiovanni and M. S. Ryoo, “Fine-grained activity recognition in baseball videos,” in CVPR Workshop on Computer Vision in Sports, 2018.
[37] Tzutalin, “Labelimg,” GitHub repository, 2015. [Online]. Available: https://github.com/tzutalin/labelImg