Data Quality And Privacy In 3D Reconstruction Based On Mobile Crowdsensing

Mobile crowdsensing is a new paradigm for collaboration derived from with development of wireless communication and sensor technology,and the widespread of mobile smart devices.It aims to leverage ubiquitous smart devices with computing and sensing abilities to collect sensing data in the environment and jointly complete sensing tasks.Crowdsensing breaks through the limitations of traditional wireless sensor networks and has several advantages,such as economic and flexible deployment,vast sensing coverage,diversified data sources,and low sensing cost.Due to these advantages,crowdsensing has received extensive attention from both industry and academia,and has been widely applied to many fields,including environmental monitoring,indoor localization,traffic monitoring,etc.3D reconstruction is a popular and challenging problem in computer vision,digital media,and computer graphics.Its main goal is to use 2D images to restore the 3D geometry of objects.Accurate 3D models have wide application prospects,such as medical diagnosis,virtual reality,and urban planning.With the development of computer hardware and the improvement of computing ability,image-based three-dimensional reconstruction has gradually become one of the current mainstream technologies.A large-scale and high-quality image dataset is crucial to the success of reconstruction.The extensive use of crowdsensing inspires us to utilize the power of crowds to collect images.However,participants will inevitably consume resources,such as network traffic,battery power,and personal time,and take the riskof visual privacy leakage while performing tasks,which makes it difficult to recruit egoistic participants to join in unprofitable sensing tasks.Meanwhile,the lack of participants will make it hard to provide enough images for 3D reconstruction.In addition,the differences in the ability and effort of participants result in uneven image quality.Low-quality images will degrade the performance of reconstruction.Besides,the system resources are limited,while the image data is in large quantities.Uploading and processing all images will require extensive resources,which is unreasonable and impractical.Especially,low-quality and redundant images may cause the waste of system resources.Therefore,we consider using monetary incentives,privacy protection,and image selection algorithms to solve above problems in crowdsensing-based 3D reconstruction.The main issues of this dissertation include designing reasonable reward distribution strategies to compensate for the cost of participants;encouraging participants to join in sensing tasks and collect high-quality images through monetary incentives;designing privacy protection methods to avoid the leakage of individual visual privacy;designing a feasible minimum image selection algorithm to achieve target coverage under resource constraints of crowdsensing systems.We analyze challenges and propose specific solutions for different goals and scenarios.Main contributions of this dissertation are summarized as follows.1.An image quality-based reward distribution strategy for crowdsourced 3D reconstructionParticipants will consume physical resources and manual effort,and take the risk of privacy leak when participate crowdsourcing tasks.Participants usually are self-interested,which makes the recruitment of participants difficult.Thus,it is necessary to have appropriate rewards or other incentives to compensate for their costs and motivate them to join in tasks.On the other hand,the gap in the personal capabilities of participants results in uneven image quality.However,a high-quality image collection is a fundamental demand for 3D reconstruction.We propose an image quality-based reward distribution mechanism,called Img Pricing,to guide participants to collect high-quality images,which is based on the idea of good work deserving good pay.We determine image quality by measuring its effective influence on 3D reconstruction.2.Image privacy protection scheme in crowdsourcing 3D reconstructionCrowdsourced images probably involve users' visual privacy.Uploading and sharing these images may expose participants to the potential leakage of personalized privacy.With the increasing of consciousness of privacy security,the leakage of personalized privacy will reduce the enthusiasm of participants and affect image collection tasks.We propose to use a pre-trained deep convolutional neural network to semantically analyze,segment and erase privacy regions in crowdsourced images.In addition,considering the information gap between segmented images and original images,which may degrade the performance of subsequent 3D reconstruction,we present an image completion network based on the idea of generative adversarial network to repair missing regions of segmented images to recover key information that is beneficial to 3D reconstruction and reduce the impact on the performance of 3D geometries.3.The minimum crowdsourced image selection under resource constraintsThe resource of crowdsourced application is limited,such as bandwidth,storage,processing capacity,and power,while the number of crowdsourced images is huge.So it is not easy and reasonable to upload and process such huge images.To balance the contradiction between resource limitation and large-scale datasets,we consider using image metadata,including geographic location and geometric information,to quantify photo utility.In addition,considering that metadata is not accuracy in actual environment,we propose the concept of photo robust utility.Besides,we formulate the minimum number of image selection problem,and present greedy algorithms based on marginal utilities to achieve coarse and refine target coverage,and reduce unnecessary consumption of resources caused by redundant or unsatisfying images.In summary,this dissertation analyzes many issues and challenges in crowdsourced 3D reconstruction,and proposes specific solutions to handle the problems including image quality,privacy protection,reward distribution,and resource constraints.The extensive experiments verify the feasibility and effectiveness of proposed schemes.