The Implemention And Research Of Lightweight Visual Odemetry For WebAR

Visual odometry is the core technology of AR application.It plays a key role in accurate pose estimation of equipments.The most popular AR devices(Microsoft Hololens,etc.)or AR frameworks(ARCore,ARKit,etc.)currently on the market use visual odometry and their variants as the basis for pose estimation.However,WebAR applications currently lack an open visual odometry solution that runs fast,consumes less resources,and can run across platforms.To get rid of this dilemma,this paper studies and implements a lightweight visual odometry,and successfully ports the algorithm to the Web platform,which ultimately improves the usability and universality of WebAR applications.To sum up,the main work of this paper is as follows:(1)In view of the lack of an open and lightweight visual odometry in the current WebAR applications,this paper comprehensively compares the three main categories of visual odometry(direct,indirect,semi-direct)and their representative algorithm frameworks(DSO,ORB-SLAM2,SVO),which depend on the running speed,third-party dependency libraries and resource consumption.The most suitable SVO algorithm framework for Web platform is selected as the basis.At the same time,in order to make up for the algorithm defects of SVO itself,a dynamic photometric compensation algorithm,a dynamic initialization algorithm and a keyframe selection algorithm are proposed as improvement measures.Tests and Analysis on TUM datasets demonstrate that the final scheme has advantages over SVO in tracking stability,initialization stability,and running speed over ORB-SLAM2 and DSO.(2)To solve the problem of how to migrate the visual odometry to the Web,this paper studies the WebAssembly-related compilation technology,compiles and optimizes OpenCV,Fast and other third-party dependency libraries,and improves the migration scheme of visual mileage calculation method to the Web.Tests on the SVO’s own dataset demonstrate the effectiveness of this migration scheme,which achieves a rate of more than 60 frames on various devices such as raspberry pie,PC,mobile phone,etc.At the same time,an additional visual inertial odometry is migrated using this scheme.Experiments on TUM-VI and EuRoC datasets demonstrate the replicability and stability of this scheme.(3)For how to design and implement a WebAR system that uses visual odometry,this paper presents a system structure that can balance data communication and data calculation.It implements various modules including initialization,world perception,world analysis,scene management and so on,which ensures fluent experience of using WebAR applications.Ultimately,it was successfully applied to the Speed Skating Pavilion of the Beijing Winter Olympic Games in 2022.Experiments on various models and systems show that the system is efficient and stable,which reduces the pressure for subsequent algorithm improvement and adaption.