Group Behavior Recognition Based On Multi-dimensional Mobile Sensory Data Fusion

Human behavior recognition is not only the key part in social science,but also important in computer science.Human behavior recognition can not only help to provide service for people,but also help to provide guidance for city management.Human behavior include individual behavior,group behavior,and community behavior.Specifically,from the aspect of the number of people,group is between individual and community.Thus,group behavior can present unique characteristics of individuals,and also obvious features of a team.Therefore,our work aims to recognize group behaviors.Comparing with individual and community,the influence among group members is more evident,and it will be reflected on group behavior.Therefore,it is necessary to capture group behavior and discover the influence.As various sensors are embedded in smartphones,smartphone can capture data related to human behavior in an un-obtrusive way.Prior works have studied the applications based on mobile phone sensed data and achieved remarkable results,especially in personal activity recognition and community tracking.However,the study on group behavior recognition is still rare.How to characterize group behavior with the limited mobile sensing data,how to depict some uncertain factors(e.g.,group member’s influence),and how to recognize group behavior effectively are challenging.We use smartphone as the sensing device,tackles these challenges,and proposes an un-obtrusive group behavior recognition method.We also made realistic environments and completed evaluation.Specifically,with the mobile sensing data collected by smartphones,we characterize group behaviors in multiple aspects,and the main work is as follows.(1)For the problem of inaccurate indoor positioning,our work proposes a moving behavior recognition method which is oriented to group-structure.It utilizes the moving information from accelerometer and position information from Wi-Fi signal,and achieves fine-grained group mobility classification and leader-follower relation recognition.For the representation of the group’s moving features,we extract relative features to depict the group’s moving status,based on the uniformity of group members’ mobility pattern.We classify group mobility into four classes,i.e.,stationary,strolling,walking,and running.Considering the difficulty and inaccuracy of indoor positioning,we identify the position relationship of group members through relative position detection.We use the motion information provided by the acceleration sensor and the position information provided by the Wi-Fi signal to describe the movement trajectories of the group members.Considering that trajectories of group members are similar with a time-lag,we propose to recognize group structure by calculating the time-lag similarity of the movement trajectories.The experiment design and data collection were completed in an office building,and the performance of group movement behavior recognition was evaluated from multiple aspects.The accuracy of leader-follower detection is 80%.The accuracy of group mobility classification is 99.5%,which is 2.5% higher than prior study.(2)The context of group interaction varies,and the representations of group interaction are not unique.Facing these characteristics,we propose a group interaction behavior recognition method,which combines travel posture and audio features.The method realizes interaction direction discovery and interactive voice detection.By observing the behavioral performance of group interactions in daily life,we find the characteristics of travel posture when the group interaction occurs,and demonstrate that the priority of the two events,i.e.,the change of travel posture,and the occurrence of voice.With multi-dimensional sensory data,we propose to recognize group interaction behavior by combining travel posture and voice communication.The experiment was completed in realistic environments,including a silent office building,a room with soft music,and a room with noisy music.The interaction detection results of the three environments are 94.4%,90.0%,and 84.2% respectively,while the interaction orientation is about 33 degree.(3)Based on the recognition of group moving behavior and interaction behavior,we propose a realistic application,i.e.,group shopping behavior recognition.As group shopping behaviors are difficult to capture and group members’ influence is difficult to quantify,our work proposes a potential buyer identification method based on the purchase intention,and an influence estimation method based on the degree of member’s interaction.Because the member’s purchase intention may be affected by her companions,we characterize the moving behavior and interaction behavior of the shopping group,and then detect different stages of group shopping.Further,on the base of these,we define the member’s purchase intention during different stages,and then identify potential buyers.As to the uncertain influence factors and their quantification,we propose two influence factors,i.e.,the similarity of members’ trajectories,and the item related interaction between group members.By utilizing multi-dimensional mobile sensory data,we quantify the factors and estimate the influence of group members.Experiments were conducted in real-life environment,and our work realized 88.9% accuracy in influence estimation.The accuracy of potential buyer recognition is 93%,which is improved by 15.4%.