A Unified Deep Learning Framework for Multi-Domain Video Summarization Using Object Detection

The topic of automated video summarization is gaining more and more significance because of the recent proliferation of multimedia information in various fields, including health, traffic monitoring, and sports analytics. Current methods of video summarization are normally targeted at particular datasets and often employ frame-level importance estimations, which restricts their capacity to be applied to different heterogeneous settings. Moreover, object detection and summarization modules tend to be used as autonomous modules, whereas temporal event modeling is loosely defined. The paper suggests a common deep learning system of multi-domain video summarization that combines object detection, temporal event modeling, and hybrid extractive-abstractive summarization. The architecture that is proposed has convolutional neural network backbones to extract features and state-of-the-art detection-based models based on YOLOv8 and Faster R-CNN to detect domain-specific objects in medical, traffic, and sports video datasets. The identified objects are arranged in time into event intervals and as a consequence, extractive video summaries and descriptions of the objects are generated in concise form. The framework can be used to benchmark across domains with the publicly available data like UA-DETRAC, CityFlow, SoccerNet, and surgical video data.