Automatic Separation of Effective Frames in the Analysis of Voluminous Videos Using Human-Object Interactions

In recent years, machine learning algorithms and deep neural networks have provided very suitable answers to solve complex problems, including the analysis of human-object relationships in videos, but one of the main problems in the operational use of these algorithms is the time-consuming processing, Especially in bulky videos. In this research, a dynamic approach is presented by combining the use of deep neural networks along with scene evaluation function with different coefficients, during which the people and objects present in the scene are identified and the effective changes are investigated, focusing on the origin of the changes. In this method, the changes of the different parts of the body are checked separately and with different coefficients. The importance of this matter is that many of the changes in the scene are not related to human-object interactions and a significant part of the processing time in the usual methods is devoted to the investigation of these changes. In this situation, most ineffective frames in the main process of the video are automatically detected and excess frames are discarded even with changes in some areas such as the background or neutral objects. The results show that this method, while reducing frames and summarizing videos with human-object relationships, helps to use deep learning algorithms for online processing of these videos.