Research And Application Of Behavior Identification Based On Geotagged Photo

With the mobile devices popularizing and developing,large amounts of multimedia sharing services,Instagram,Flickr and others,emerge on the internet.The photo-sharing Services has led to the explosion of Geotagged Photos,which have text tags,timestamp and geographical coordinate.However,due to the serious fragmentation of these photos,users will not sort and classify them,which will waste a lot of time during the process of browsing and finding.Therefore,how to systematically manage and classify a large number of photos is a question.Meanwhile,the rules of the users' behaviors can be inferred from the relationship between the users' behaviors and the significant locations.It is one of the heated issues in the area of social computing.To solve these problems,based on the connection between location and behavior,this paper chooses the individual discrete Geotagged Photo as experimental data,which is collected by the mobile terminal App overs 20 months,and the App was developed by the author's laboratory.According to the location,time stamp,text descriptions and the geographically uneven distribution characteristics,this paper proposes a top-down step by step refinement behavior recognition method.First of all,according to the actual address of the each image,the visited area are divided into several regions and the areas with less data are filtered;Then,selecting the areas with high density profile from the classified areas,we use grid algorithm to find activity concentration areas,which will be defined as significant activity areas;Then,according to the location distribution density of the image in the significant activity areas,we used a clustering algorithm based on distance and density,and the each cluster will be defined as a significant location.Correspondingly,the semantic descriptions of the image are extracted from each cluster.We will extract keywords based on word frequency,and using the keywords to label the significant location.Compare the similarities between the marked words of each cluster with the keywords in non-cluster location.If the degree of similarity is high,this location is additionally identified as a significant location.On the contrary,the insignificant location at a low degree of similarity.Finally,extracting the features of the image within each cluster identified as significant location,we will use using clustering algorithm based on the distance to gather these image in one significant location.Mark the behavior of the cluster using the keywords which are extracted from each cluster based on the word frequency and build the index according to the results obtained by a multi-layer identification.Thus,the data is sorted,and the important location and behavior can be recognized.In this paper,we obtained 1239 valid data from multiple cities,Shenyang,Siping,Beijing,Yingkou.During the experiment,the optimization and combination of the parameters at each stage are taken into consideration in this paper.The clustering results of the significant locations evaluated by using multiple clustering indexes.The significant location identified by the system are corrected based on the significant locations marked by the users.The behaviors are identified after the important locations are corrected.The evaluating process is conducted by comparing the marked and the identified behaviors,and the result show that the similarities between the automatically identified behaviors and the actually marked behaviors are satisfying.