Human Posture And Action-based Video Retrieval

With the development of Internet and the capacity of hardware,the number of video data increases sharply.Therefore,the establishment of effective video data management methods become more and more urgent needs.Compared with picture and text,video holds more information,video retrieval is more challenging.Video retrieval includes three parts: feature extraction,similarity sorting and re-ranking.This paper mainly studies the graph-based reranking method,we propose the shortest edge clustering algorithm based on TTNG method,and apply it to the multi-feature fusion re-ranking method.Based on this method,a human action-based video retrieval framework is proposed.Firstly,this paper proposes a new human action data set: RGB-D human action-emotion data set.This data set is divided into seven categories: happy,anger,fear,disgust,sadness,surprise and normal,with a total of 4224 samples.Compared with other commonly used human action data sets,this data set has the characteristics of pure background,fixed camera and light,multi-angle synchronization,RGB and depth video synchronization,only setting the situation but not setting the actions.In the process of video retrieval,the wrong results are mainly divided into two categories: outliers and samples that are close to the query sample but belong to other categories.The TTNG algorithm proposed by Muxin Sun can solve the first category well.In order to solve the second category,we propose the shortest edge clustering algorithm.The validity of our algorithm is proved by experiments.In order to further improve the retrieval effect,we apply this algorithm to the multi-feature re-ranking.Based on this method,a human action-based video retrieval framework is proposed.In addition,in order to improve the time performance of this framework,we propose two improvement methods for the shortest edge clustering algorithm:(1)Clustering on the neighbors of query q only.(2)In order to avoid updating the distance between sets for many times,we propose a clustering method similar to union-find set.After the improvement,the time performance of our video retrieval framework has been improved obviously.