Research On High-quality Data Collection For Mobile Crowdsensing

Mobile CrowdSensing(MCS)leverages people's smart devices to efficiently accomplish many complex sensing tasks of different domains,such as transportation,society,and environment.Through the mobile internet,people intentionally or unintentionally collaborate,which brings MCS the opportunity to allocate tasks and gather data.The rapid development of sensors has brought in various types of MCS data.Besides the traditional numerical sensor data,more and more MCS data are in the form of audios,images,and texts.Photos are widely used by MCS.Non-redundant photos that can meet the requirements of the task are high-quality MCS data.MCS data collection involves the bilateral interests of data providers(workers)and data requesters.The research aiming to collect high-quality MCS data in an efficient and low-cost way faces the following two challenges: i)Using rational task allocation to assure sufficient data sources;ii)Using data selection to improve the quality of data set.Based on the challenges,this dissertation will propose methods to improve the efficiency of task allocation,data quality measurement,and data aggregation.The contributions of this dissertation are as follows:(1)Proposed a task allocation algorithm that is applicable to MCS task allocation with spatial constraints,and a multi-facet task allocating optimization algorithm so as to improve the allocation rate and assure sufficient data sources.On one hand,to balance the sensing cost and task allocation rate,a task allocation algorithm based on the spatial relations between the task scene and people's travel plans was proposed.This will improve the task allocation rate and assure sufficient raw data constrained by the sensing cost.On the other hand,according to the physical features and data collection requirements of the sensing objects,a multi-facet sensing task allocation algorithm was proposed,and this will break the task into subtasks of sensing the object from various facets.By using this optimal task breaking method,the allocation rate will be further improved.(2)Proposed a generic MCS photo collection framework and a data selection algorithm of the photo stream for different tasks so that only high-quality data will be collected at a low cost.MCS tasks' data are heterogeneous and the MCS data will be streaming to the data center,so this dissertation proposed a generic task model and the pyramid tree model in order to set constraints of collecting data for different tasks.Since the MCS data are streaming to the data center,this dissertation also proposed a pyramid-tree-based data clustering method,which can efficiently and precisely obtain the near-optimal result of the data selection.(3)Proposed a method to forward MCS data based on the tree-fusion in the opportunistic network in order to save the cost of collecting high-quality data.Forwarding MCS data by using epidemic routing requires wider bandwidth and costs more traffic in the opportunistic network.To improve the quality of data collection by only efficiently forwarding the valuable data from participants,this dissertation proposed a data set fusion method based on a PTree.This method combines PTrees carried by different participants and then determines which data should be treated as high-quality and should be forwarded to others.Through this method,only high-quality data will be forwarded to participants.Therefore,the communication and storage overhead of data collection are largely reduced.(4)Developed an MCS-based event-sensing system that collects high-quality photos of the event in real time and proposed a mobility-based data selection method.This system collects MCS data contributed by witnesses to learn the progress of the event.Through leveraging both the relationship between the crowd's mobility and the photo stream's instability and the relationship between the crowd's mobility and the event's progress,this dissertation proposed a photo stream segmentation method based on the sub-event detection and proposed a key photo selection method based on sub-event information entropy.These methods are utilized to divide the photo stream into segments that can correspond to different sub-events one by one.Based on these segments,we can get a high-quality photo set that has less redundant photos as well as can widely cover the event.In conclusion,this dissertation proposed a series of models and methods to sample and collect high-quality photos for mobile crowdsensing.Through theoretical analysis,experiments and an application system,these models and methods are proved to be applicable and can be used as theoretical and technical references for MCS applications in different domains.