針對人體骨架座標缺失值之填補法比較

Abstract

科技的快速發展使人人都能夠輕易地擁有大量的數位影像資料，影像辨識的需求也隨之增加，藉由影像辨識技術能夠取代過往需要倚靠人力長時間監測的工作，提升工作效率。 人體辨識即為影像辨識中相當重要的一部分，大多藉由骨架資料來觀測人類的動作。取得骨架資料的方式很多，例如：穿戴式裝置或是感測裝置等，有些低成本的感測裝置雖然容易取得，但是產生的資料容易會有雜訊、離群值或缺失值，大量的缺失值很可能會影響後續研究。 本研究將依照實驗設計的原則來模擬並產生缺失值，用不同方法對缺失值進行填補，並考慮骨架長度的限制對填補後的結果進行修正，最後針對不同情況下的缺失值給予填補方法建議。

The rapid development of technology allows everyone to easily have a large amount of digital image data and the demand for image recognition has also increased. With image recognition technology, we can replace the work that used to require long-term manual monitoring and improve efficiency. Human recognition is an important part of image recognition. It mostly uses skeleton data to observe human activities. There are many ways to obtain skeleton data, such as wearable devices or sensing devices. Although some low-cost sensing devices are easy to obtain, the generated data is likely to have noise, outliers or missing values. A large amount of missing value might affect subsequent studies. In this study, we will simulate and generate miss values according to the principles of experimental design. Then use different methods to impute missing values and modify the result after imputation by considering the limitation of the length of the skeleton. And lastly, we will give suggestions of missing values imputation according to different situations.