- 學位論文
一個以三軸加速規訊號為基礎的自動化游泳姿勢辨識系統
An Automated Swimming Style Classification System Based on the Signals from a Three-axis Accelerometer
摘要
隨著醫學發展的不斷進步，民眾對自身健康的問題越來越重視，加上體育休閒運動的推動，運動人口也持續地成長。游泳運動是有益身心健康的全身性休閒運動，能使全身均衡的發展，且適合所有年齡層民眾的參與。因此行政院體育委員會從2010年起正式推動「泳起來！專案」，希望培育專業人才以及改善全國游泳環境和提升全民游泳的能力，至於在業界方面，則希望開發出一套可以提供使用者自我運動管理的系統。 本研究探討如何利用三軸加速度感應器來進行游泳姿勢的辨識，達到評估各姿勢間游泳運動量的目的。在特徵提取部分，本研究經過許多嘗試，最後保留最好的兩個特徵，分別是視窗內局部振幅最大值與最小值的差以及振幅值大於某個指定值的樣本個數。在辨識器方面，本研究利用最小距離分類器以及倒傳遞類神經網路來做姿勢的辨識，此兩種方法可以有效的區分出四種游泳姿勢(自由式、蛙式、仰式和蝶式)。實驗結果顯示，在訓練集方面，自由式的辨識成功率為100%，蛙式為99.20%，仰式為100%，蝶式為98.75%；而在測試集方面，自由式的辨識成功率為97.12%，蛙式為98.45%，仰式與蝶式為100%。 未來還可以增加雲端系統，將使用者所記錄的泳姿資訊上傳到雲端系統，就可進行長期泳姿資訊的記錄，有助於使用者評估自身各姿勢間游泳運動量的程度以及提供相關資訊的回饋。
並列摘要
As medical development keeps improving, people are more and more concerned about their own health. Furthermore, the promotion of leisure sports makes the population of sports to grow continuously. Swimming is systemic of leisure movement; beneficial for physical and mental health; and enables the balance of body development. Also it is suitable for all ages of people to participate. Therefore the National Council on Physical Fitness and Sports of Executive Yuan officially promotes the project called “swim up!” from 2011. They want to cultivate professionals, improve the national swimming environment, and enhance people’s ability of swimming. As from the industry aspect, they want to develop a system which can provide the user self-management of exercise. How to apply a three-axis acceleration sensor in order to classify swimming styles and evaluate the physical activity of swimming are discussed in this study. In the feature extraction part, the two best features, namely the amplitude difference between the highest and the lowest sample values in a local window and the number of samples whose amplitudes are greater than a specified number, are adopted after the trials of many possible features. In the classification part, the minimum distance classifier and back propagation neural network are used to swimming style classification. Both methods can be effective in distinguishing four swimming styles (freestyle stroke, breaststroke, backstroke, butterfly). For the training set, the experiment results show 100% successfully identification on freestyle stroke, 99.2% successfully identification on breaststroke, 100% successfully identification on backstroke and 98.75% successfully identification on butterfly. For the testing set, the experiment results give 97.12% successfully identification on freestyle stroke, 98.45% successfully identification on breaststroke, 100% successfully identification on backstroke and butterfly. In the future, the proposed system can also include a cloud system which can upload user’s recording information and long-tern recording of swimming information. It can help users to estimate their own amount of exercise regarding the extent of swimming among four styles and provide relevant feedback information.