Keyframe Extraction Approaches for Videos under Multimodal Scenarios: A Survey

Abstract

Video keyframe extraction technology, as a core component of video summarization, retrieval, and content analysis, continues to receive widespread attention in the field of computer vision. With the rapid development of diverse applications such as smart cities, autonomous driving, and human-computer interaction, the sources and scenarios of video data have become increasingly complex, placing higher demands on keyframe extraction techniques. In dynamic scenes, factors such as intense environmental interference and highly variable motion of subjects often cause traditional extraction methods to face challenges like insufficient robustness and limited extraction accuracy. To address these issues, numerous intelligent extraction algorithms based on deep learning have emerged in recent years, significantly enhancing extraction performance in dynamic environments. By systematically reviewing keyframe extraction methods suitable for various dynamic scenes including surveillance, sports, and gesture analysis, and summarizing mainstream algorithmic models based on attention mechanisms, temporal modeling, and reinforcement learning, this paper analyzes their core concepts, advantages, and limitations. It concludes with experimental findings and offers personal insights, aiming to provide a clear reference framework and development direction for future research on efficient and robust video parsing technologies in dynamic scenes.