Research On Indoor Visual Positioning Method Based On Adaptive Threshold RANSA

With the rise of intelligent mobile terminals and the Internet of Everything,high-precision positioning and navigation technologies have progressively assumed a prominent position in the information technology sector and other industries.Traditional RF positioning technologies such as WIFI,Bluetooth,and Zig Bee are highly dependent on signal source coverage,and signal propagation quality is readily affected by indoor walls,making it difficult to achieve satisfactory positioning results.Consequently,indoor visual positioning techniques that utilize the visual image characteristics of indoor environments have a significant research value.In addition,as an emerging indoor positioning method,indoor visual positioning is not constrained by infrastructure and has additional benefits in terms of low cost,simplicity of understanding,and environmental adaptability.Visual localization methods that rely on the environment’s own visual features have to establish the visual a priori map of the localized scene in the offline stage to store the complete image features as well as the precise geographic location,in the same way,that traditional RF localization techniques must build a signal strength map in advance.In the process of establishing an offline visual map,however,an improper selection of fixed sampling spacing can easily result in redundant image samples or missing visual features from some scenes.In this study,we proposed a construction method based on the similarity of pre-sampled images,develop a similarity determination model to determine the optimal sampling spacing for indoor environments,so as to maintain the integrity of the visual features of indoor scenes while effectively controlling the database size.In the online phase of visual positioning,image traversal and retrieval in the visual map have to be completed based on the user-input query image to retrieve database images similar to its content.However,inferior image retrieval accuracy and greater overhead time can significantly degrade moving target localization precision.This study proposed a deep hashing-based image retrieval method and increases the efficacy of image retrieval by improving the optimizer,model parameters,and loss function.In this study,image retrieval performance will be evaluated using a public dataset and the visual map,and retrieval model accuracy will be evaluated using three metrics: mean average precision,precision-recall curve,and Top N precision curve.Finally,retrieval model time overhead will be evaluated by calculating image encoding time.After obtaining the image query results,the visual localization method completes pose estimation by the epipolar geometry relationship between images.However,considering that the Random Sample Consensus(RANSAC)algorithm with fixed thresholds cannot provide accurate positional estimation results in the face of complex and variable indoor environments and has poor environmental adaptability,this study adopted an improved adaptive thresholding RANSAC(Optimized Random Sample Consensus,ORSA)algorithm.In addition,to handle the error matching correspondence that still exists in the estimated inliers set,this research proposed a rejection method based on the slope density of matching correspondence to reoptimize the inliers set and reduce the solution error of the fundamental matrix,thereby improving the final positioning precision.The results of simulation experiments and real-world applications demonstrate that the proposed visual localization method achieves higher precision and outperforms the fixed-threshold RANSAC algorithm in terms of inliers rates,illumination conditions,and the number of feature points,and is of high engineering application value.