Research And Realization Of The Improvements Of Localization Methods In WebAR Navigation Systems

In the field of WebAR,navigation is a common use case.During the process of navigation,locating systems that utilizes visual matching usually employs traditional or deep learning interest point extraction algorithms for the construction of point cloud map;while the locating is done via collecting real-time images,and the matching of interest points in the point cloud.The present method is,although is capable of completing the task of navigation,only takes single images into consideration during the construction of point cloud map,failing to acknowledge that in the real life workflow process,WebAR navigation tasks could obtain images in the form of video stream,causing the algorithm to overlook past information,resulting in a less than satisfactory point cloud map;in the meantime,considering the latency caused by matching,present WebAR navigation systems based on visual information still has plenty of room for improvement.By extracting interest points with higher quality to construct better point cloud maps,locating results could achieve higher accuracy,while simultaneously running a light-weighted locating algorithm on Web,present visual WebAR systems could achieve higher accuracy and better speed.For interest point extraction,this paper employs video stream as input,fully utilizing the temporal information that users acquire on Web devices.By comparing interest point likeness in adjacent images in video sequences during the model training process,corresponding loss functions are designed to optimize model parameters,achieving higher stability in the output interest points.Higher stability interest points could form higher quality point clouds that can provide more accurate locating results,improving the overall accuracy of the locating service.As for the lightweight locating algorithm,this paper does fast estimations of the user’s current location through directly acquiring the readings of the inertial sensors on the Web device,which form a pedestrian dead reckoning system.Pedestrian dead reckoning forms a loosely coupled system with the cloud,in the event of network fluctuations that is common in indoor aeras,it could maintain basic WebAR navigation system functions.Meanwhile,this algorithm utilizes the precise location results from the cloud,calibrating the algorithm during runtime,and could keep the error as small as possible,which means the system could maintain high accuracy while greatly improving the response time.Based on above work,this paper proposes a WebAR navigation system,this system could locate the user’s precise location with high accuracy and fast response time,even in complex environments like indoor aeras.For this purpose,this paper designed a system architecture that employs endcloud collaborations,while the Web side handles lightweight tasks,the cloud side processes heavy computation loads,forming a complete system together.In the meantime,the cloud part of the system is based on docker,which means it is quite suitable for future relocation,deployment,or expansion related tasks.