標題: 基於類神經網路反回饋機制的可適性羽球擊球分類
Adaptive Badminton Stroke Classification by ANN Backward Propagation
作者: 王致皓
易志偉
Wang, Chih-Hao
Yi, Chih-Wei
網路工程研究所
關鍵字: 羽球;智慧球拍;穿戴式設備;球種分類;可適性模型;神經網路;Wearable Sensor;Machine Learning;Stroke Classification;Badminton
公開日期: 2017
摘要: 我們設計了智慧羽球拍,並利用機器學習技術,分析Random Forest,SMO(Polynomial Kernel, RBF Kernel)和Naïve Bayes四種分類器對於球種分類的準確度。在以前的研究中,發現了在一般模型和個人模型之間有個明顯的差距,約15-20%。然而,以前的機器學習技術是批量處理,沒有從一般模型到個人模型微調的方法。在本研究中,我們想基於神經網路和應用反向傳播,有機會將一般模型個人化,藉由使用即時的使用者資料來微調模型,可以讓一般模型從80%的準確度提升至93%的準確度,平均只要5組資料便能提升10%的效能。 另一方面,我們開發一個手機應用程式來實現我們的想法。它基於穿戴式設備和行動計算技術,讓使用者可以透過我們的智慧球拍建立自我訓練和比賽時的擊球記錄,通過錄影輔助可以靈活地標記球種,讓使用者標記優良測資和更正錯誤的資料,透過兩個手機程序互相連接交換資料,實現對戰計分的功能,再結合雲端服務,可以長時間地分析個人的打球風格和練習記錄。藉由發現選手的弱點來提升選手的羽毛球能力。
We have developed a smart badminton racket prototype with machine learning labeled technique to do automatic stroke type classification by Random Forest, SMO and Naïve Bayes. In the previous works, it shown that there is a gap between general model and personal model with accuracy 80% to 95% in average. However, the previous machine learning technique such as Random Forest, SMO and Naïve Bayes is batch processing and don’t have fine-tune method from general to personal model. In this work, we would like to design based on Neural Network and applied back-propagation to have the chance to modify general to personal. On the other hand, we develop an App to realize our ideas. It is based on wearable devices and mobile computing to log stroke types in real time. Through the video program can label stroke types in flexible and automatic. Let two mobile phone programs communicate with each other to battle a game and score points. By combining with cloud service, it could be used to analyze personal style of play and game record in a long period. To improve the badminton rally ability of players by finding out the weakness of them.
URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070456524
http://hdl.handle.net/11536/142669
Appears in Collections:Thesis