| With the development of the Internet and multimedia technology,dynamic digital resources represented by videos have gradually replaced static digital resources such as text and pictures,and become the mainstream mode for people to record and share their lives.Video content understanding has gradually become a research hotspot in the field of computer applications,because it has an urgent need in practical applications.Social relationships are an important part of people’s daily life.Understanding the social relationships of people will help downstream tasks such as accurate recommendation,behavior prediction,and video question answering.In recent years,identifying social relationships of characters in videos and generating knowledge graphs of character relationships has become a key issue in computer vision and multimedia understanding.At the same time,video data has a large amount of information,takes a long time to process,and takes up a lot of hardware resources.Therefore,the operating performance of algorithms related to video content understanding and the consumption of computing resources have become the key bottlenecks affecting its implementation.The main research and development work can be summarized as follows:1)Firstly,a video character representation method is proposed,based on visual information and timing information,video characters are represented by multiple entities,and a social relationship knowledge graph is generated.Secondly,a multi-entity relationship reasoning framework MRR is proposed,which collects,disseminates,and aggregates information from the neighborhood of entities based on the attention mechanism,and then predicts the relationship between characters according to the context information and the updated knowledge graph.Finally,the comparison experiments and ablation experiments completed on the MovieGraphs dataset prove the effectiveness and advancement of the MRR framework.2)An end-to-end social relationship knowledge graph construction algorithm is proposed.In view of the large amount of video data,long processing steps,and different storage and computing resources required for each step,the speed of the algorithm and the efficiency of resource usage are improved by modularizing the processing steps,parallelizing the module operation,and dynamically allocating resources.The performance experiment proves that the algorithm is more efficient than the general serial algorithm.3)Develop and build a video character social relationship knowledge graph construction system,which provides functions such as video data management,video social relationship knowledge graph construction,and result visualization,which is helpful for ordinary users to directly understand the achievements in this field.Researchers can also analyze data results through this system,which is helpful for the research of downstream tasks.In addition,the modular design is also easy to integrate algorithms,which is conducive to the implementation of algorithms related to video content understanding. |