Automatic Metadata Generation Through Analysis of Narration Within Instructional Videos

摘要

Current activity recognition based assistive living solutions have adopted relatively rigid models of inhabitant activities. These solutions have some deficiencies associated with the use of these models. To address this, a goal-oriented solution has been proposed. In a goal-oriented solution, goal models offer a method of flexibly modelling inhabitant activity. The flexibility of these goal models can dynamically produce a large number of varying action plans that may be used to guide inhabitants. In order to provide illustrative, video-based, instruction for these numerous actions plans, a number of video clips would need to be associated with each variation. To address this, rich metadata may be used to automatically match appropriate video clips from a video repository to each specific, dynamically generated, activity plan. This study introduces a mechanism of automatically generating suitable rich metadata representing actions depicted within video clips to facilitate such video matching. This performance of this mechanism was evaluated using eighteen video files; during this evaluation metadata was automatically generated with a high level of accuracy.