One of the fundamental steps for video processing is to decompose a video into meaningful segments. In previous methods, reliable result for the videos recorded under non-stationary environment could not be produced. Moreover, backward-walk slide transition that similar slides are repeated also cannot be detected. We propose a lecture video segmentation method which is available for various type videos and consider the lecturer behavior by using SIFT, adaptive threshold and graph based detection model. The way to know whether slide to slide transition happens or not is made up two main steps. First is similarity computation for which the slide region is detected then similarity between slides is computed by using SIFT. In second step, the graph based transition detection model will be constructed based on the similarity. At that time, threshold is adjusted adaptively according to characteristic of relevant cacheable segment. The slide transition point can be gotten from graph-based transition detection model and adaptive threshold. The experiments performed on various video types show the effectiveness of ours.