Due to the development of AI technology, in order to recognize and understand the surrounding environment, speech and visual abilities, language understanding, action planning, and prediction are required. Therefore, effective video captioning continu...
Due to the development of AI technology, in order to recognize and understand the surrounding environment, speech and visual abilities, language understanding, action planning, and prediction are required. Therefore, effective video captioning continues to require advanced capabilities to predict the behavior of nearby objects and generate appropriate captions for future scenarios. This has a limitation in that it relies on user behavior for prediction, which limits the information that can be provided to the user. For this, it is necessary to develop a technique that provides behavioral information to users in various environments. The method proposed in this thesis describes visual content by analyzing the behavior and characteristics of objects in the video, and generates captions for the input video by combining global and local characteristics of visual features. In this thesis, I verify and evaluate the performance of an effective video caption generation method by comparing the behavior and attributes of objects in the video with convolutional neural network methods such as ResNet and ViT, which are individually pre-trained on the visual features of the input video.