Multiple Viewpoint Classroom Recording System and Study on Automatic Lecture Video Summarization

[[abstract]] 近年來數位學習的發展已成為未來教學模式的趨勢，數位學習在科技蓬勃的發展下，包括錄製設備、硬體容量的提升與網路的傳播，讓教學影片有高解析畫面做紀錄並快速傳播學習的內容。因為教學影片完整記錄整個教學現場，除了供給學生作為學習外，更可以觀察課程中導師與學員反應，作為導師間互相交流心得與學習教材。但教學影片內容除了依賴導師事前課程規劃外，經過編輯擷取摘要的教學影片有更好的學習效果。 本論文為了達成更完整記錄課堂的目的與研究資料的收集，嘗試使用各種攝影設備與觀察分析課堂環境，挑選出符合課堂環境的設備與最佳錄製方法，並建構出一套可以自由切換視角與不同畫面類型的播放系統並互相討論交流觀課心得的網路頁面平台。本論文亦發展影片自動摘要的方法，讓觀看者進入完整課程前能先觀看摘要影片，了解課程導師的上課情境與風格，方便找出觀看者覺得可能有興趣的教學影片。 自動影片摘要的研究先收集 94 筆課程紀錄影片資料集，各個類型課程都是以連續且沒有編輯過的完整課堂紀錄影片為主。研究分為三階段，第一個階段針對課程影片做影像校正，校正影像帶入由ETH Zurich, M. Gygli 團隊 [1] 提出一種片段化的以供影片摘要的方法，第二個階段從這些片段中分析可能有興趣或重要的片段特徵，透過觀察各個課程類型提出三種特徵：光暗亮度、聲音音量與畫面動態運動幅度，第三階段經過片段分析後綜合三項特徵做出評分推薦適合的片段，綜合推薦片段得出影片摘要。 得出影片摘要後，除了客觀分析影片摘要是否合適，還製作了一份主觀性問卷，問卷主要目的是讓測試者挑選哪一種影片摘要的方式得到有興趣或是重要的片段內容，並測試適合的摘要影片長度，最後從問卷回饋中得出影片摘要結果評估與改善的方向。[[abstract]]Recently, E-learning has become a trend of future teaching. Because the science and technology flourished in recording equipment, hardware and communication network capacity, the lecture videos have higher recording resolution and the content can be spread more quickly. Through a more complete recording of the whole teaching site, it can be used for collaborative lesson preparation to observe the interaction between the instructors and the students. A good lecture video not only rely on the instructor's prior course planning, but a proper edited video summarization can also improve the efficiency of learning. In order to achieve a more complete record of the classroom and for the collection of research materials, this thesis have used a variety of photographic equipment and analyzed the classroom environments to choose the best recording method suitable for the classroom environment. After the recording, we have developed a web platform that can playback freely among lecture videos of the same classroom in different viewing angles and recording types, and provide a function for discussion and exchange ideas of lesson preparation. In this thesis, we also developed automatic video summarization algorithms to let viewers watch the summary movie before entering the full course. The summary videos help the viewers to understand the style of the course instructors, and to choose lecture videos that they may be interested in. For the study of automatic video summarization, the data set was collected comprised of a total of 94 class record videos, including each type of class recorded continuously and unedited as a complete classroom lecture video. Our research is divided into three phases. In the first phrase, the image correction is performed for the lecture video, and the corrected image is brought into a fragmentation method for the video summarization proposed by the ETH Zurich, M. Gygli [1] team. In the second phrase, we analyzes the features of the fragments that may be of interest or importance from these segments, and proposes three characteristics by observing each course type: brightness, sound volume and dynamic motion range of the video. In the third phrase, after the segment analysis, the three features are combined to make a score suitable for the recommended segment, and the recommended segment is combined to obtain a video summarization. After obtaining the summary video, in addition to the analysis of the content of the video, we produced a subjective questionnaire. The main purpose of the questionnaire is to allow the subject to choose which of the summarized videos has picked the interesting or important fragments and which length of summarized video is proper. Finally, the evaluation and future improvement of the video summarization are obtained from the questionnaire feedback.[[note]]碩