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

[[abstract]]近年來影片式的數位學習是非常熱門的一種學習方式，但教學者為了做出好的教學影片，往往需要耗費相當大量的時間來對教學影片進行剪輯，剪輯的時間成本以及剪輯技術成為了教學者的負擔及入行門檻；而且大多數的影片教學僅用一隻固定式攝影機對著講師拍攝，在某些場合下會無法滿足學習者的要求，例如：講座式教學的授課內容包含有講師授課及小組討論，那僅用單一固定式攝影機就會無法拍攝到小組討論的部分。 所以本論文採用全景及多隻廣角攝影機來完整記錄整間教室的授課內容，並建構一個讓使用者能自由切換多種視角或不同類型(全景、廣角)影片、討論觀課心得的整合平台，也提供一種教學影片自動分段的方法，讓教學者省去傳統人工剪輯的大量時間；自動分段後會產生數個分段點，學習者可以利用分段點來快速的了解該場次的影片內容。 自動分段方法是先從影片取出影像來進行魚眼校正、轉灰階影像以及影像增強，再利用Mean Square Error來比較前後兩張影像的差值，當差值大於門檻值則視為分段點。講座式的教學影片是使用對準教室前方的影片，並僅對投影布幕來進行 Mean Square Error偵測；活動式的教學影片則是使用在教室正中央由上往下的影片，對完整的影像來進行 Mean Square Error偵測。 自動分段實驗結果評估採用 Precision、Recall以及 F1-Measure來評估，講座式的教學影片實驗結果為 Precision 97.1%、Recall 87%、F1-Measure 91.7%，活動式的教學影片實驗結果為 Precision 76.9%、Recall 90.9%、F1-Measure 83.3%，從實驗結果來看講座式的教學影片成果較佳，而活動式的教學影片尚有改進的空間。[[abstract]]In recent years, the e-Learning is a very popular learning method, but the lecturer often spends a considerable amount of time to edit the teaching video. The time cost of editing and editing skill become the burden of the lecturer and the threshold for entry. In addition, most of the teaching videos were recorded using only a fixed camera at lecturer position. In some cases, it will not meet the learner's requirements. For example, lectures video including the lecturer teaching and group discussions is unable to be captured with a single fixed camera because the part of the group discussion is missed. Therefore, we uses panoramic and wide-angle cameras to completely record the lecture content of the whole classroom, and construct an integrated platform that allows users to freely switch between multiple viewpoint or different types of (panoramic, wide-angle) videos, and discuss about lecture experience. We provide a method for automatically segmenting teaching videos, make the lecturer to save a lot of time for editing videos. After segmentation automatically, several segmentation points are generated, and the learner can use the segmentation points to quickly understand the content of the video of the session. The automatic segmentation method first extracts images from the video for fisheye calibrate, then convert them to grayscale images and apply image enhancement. Mean Square Error is used to compare the difference between consecutive two images. When the difference is greater than the threshold, it is regarded as a segmentation point. For lecture-based teaching classroom, the shot at front of the classroom is applied for Mean Square Error detection on area of the projection screen. For active teaching classroom, we use the video on the top in the center of the classroom, and the whole image is used for Mean Square Error detection. The results of the automatic segmentation experiment were evaluated by Precision, Recall and F1-Measure. For lecture-based teaching classroom, the results were Precision 97.1%, Recall 87%, F1-Measure 91.7%. For the active teaching classroom, the experiment results were Precision 76.9%. Recall 90.9%, F1-Measure 83.3%. From the experimental results, the lecture-based teaching classroom has better segmentation performance, and the active teaching classroom still has room for improvement.[[note]]碩